{
  "description": "Existing aggregate reports only; no predictions, prompts or hosted Jev results. No new evaluation was run for this post. Intervals are group-clustered bootstrap CI95, conditional on each checkpoint, not across seeds.",
  "source_repository": "https://github.com/nullsilver-labs/alpha-sys-1",
  "source_files": {
    "runs/cmp/metrics.json": {
      "sha256": "3006346d440031e5d6ced935d11c7bf95f369c9fe6d163c5e7504fb6c272144b"
    },
    "runs/p2/gate-7env-450M.json": {
      "sha256": "a9200677ecf2dfb9e13dc364bd188a4ab0034f9b1c16c5b92ab23dc6fb9cc1d5"
    },
    "runs/p2/gate-7env-1.6B.json": {
      "sha256": "24dadd794fd3954cb19e1c6e05a67875ef0363e35ce5822f1c2f7175cfa73aa5"
    },
    "runs/p2/gate-7env-3B.json": {
      "sha256": "72139e3384c641c7abd902083fbf162e9dc70d9838ac637d24efa2003594f444"
    }
  },
  "release_seeds": {
    "450M": 1,
    "1.6B": 1,
    "3B": 1
  },
  "release_revisions": {
    "450M": "260919",
    "1.6B": "260919",
    "3B": "260920"
  },
  "trained_environment_baselines": {
    "450M": {
      "mcq": "LFM2.5-VL-450M +T",
      "chaosnli": "zs-LFM2.5-VL-450M-prefill +T",
      "civilcomments": "zs-LFM2.5-VL-450M-prefill +T",
      "stsb": "zs-LFM2.5-VL-450M-prefill +T",
      "folktables": "zs-LFM2.5-VL-450M-prefill +T"
    },
    "1.6B": {
      "mcq": "zs-LFM2.5-VL-1.6B-prefill +T",
      "chaosnli": "zs-LFM2.5-VL-1.6B-prefill +T",
      "civilcomments": "LFM2.5-VL-1.6B +T",
      "stsb": "zs-LFM2.5-VL-1.6B-prefill +T",
      "folktables": "LFM2.5-VL-1.6B +T"
    },
    "3B": {
      "mcq": "zs-LFM2.5-VL-3B-prefill +T",
      "chaosnli": "zs-LFM2.5-VL-3B-prefill +T",
      "civilcomments": "zs-LFM2.5-VL-3B-prefill +T",
      "stsb": "LFM2.5-VL-3B +T",
      "folktables": "LFM2.5-VL-3B +T"
    }
  },
  "unseen_baseline_policy": "Fixed first-assistant-token LFM readout + dev-fitted temperature. Not the lowest test CE of two readouts.",
  "metrics": [
    {
      "env": "boolq",
      "slice": "all",
      "system": "base rate",
      "metrics": {
        "n": {
          "value": 8129.0,
          "lo": 7984.475,
          "hi": 8270.575
        },
        "ce": {
          "value": 0.6646544945813517,
          "lo": 0.6585364928682439,
          "hi": 0.6720828659936188
        },
        "kl": {
          "value": 0.6646544945813517,
          "lo": 0.6585364928682439,
          "hi": 0.6720828659936188
        },
        "jsd": {
          "value": 0.20396171787752962,
          "lo": 0.2023067701872962,
          "hi": 0.20597112661708397
        },
        "brier": {
          "value": 0.4717530550578688,
          "lo": 0.4657881407089929,
          "hi": 0.47899555045156544
        },
        "acc": {
          "value": 0.6198794439660474,
          "lo": 0.6065282404082177,
          "hi": 0.6308754862619356
        },
        "conf": {
          "value": 0.635615028306742,
          "lo": 0.635615028306742,
          "hi": 0.6356150283067422
        },
        "ece": {
          "value": 0.015735584340694553,
          "lo": 0.004739542044806519,
          "hi": 0.029086787898524334
        },
        "reliability": {
          "value": 0.0002476086145431116,
          "lo": 2.2486951592956432e-05,
          "hi": 0.0008460416293660667
        },
        "resolution": {
          "value": 0.0,
          "lo": 0.0,
          "hi": 0.0
        },
        "uncertainty": {
          "value": 0.2356289189143913,
          "lo": 0.23287158340290345,
          "hi": 0.23865173359641664
        },
        "auroc": {
          "value": 0.5,
          "lo": 0.5,
          "hi": 0.5
        },
        "acc@80": {
          "value": 0.6226357065969552,
          "lo": 0.6052843777392773,
          "hi": 0.6335721802831055
        }
      }
    },
    {
      "env": "boolq",
      "slice": "all",
      "system": "laya-0.1.6",
      "metrics": {
        "n": {
          "value": 8129.0,
          "lo": 7984.475,
          "hi": 8270.575
        },
        "ce": {
          "value": 0.16732241000325002,
          "lo": 0.15388610376208772,
          "hi": 0.1802310830968935
        },
        "kl": {
          "value": 0.16732241000325002,
          "lo": 0.15388610376208772,
          "hi": 0.1802310830968935
        },
        "jsd": {
          "value": 0.041991452965761206,
          "lo": 0.039388778184086294,
          "hi": 0.044628730457879705
        },
        "brier": {
          "value": 0.08883623468224963,
          "lo": 0.08079539790469732,
          "hi": 0.09667981589340952
        },
        "acc": {
          "value": 0.9453807356378399,
          "lo": 0.9399843400447427,
          "hi": 0.9504083064841119
        },
        "conf": {
          "value": 0.9455707663465563,
          "lo": 0.9437346472223621,
          "hi": 0.9474041161979702
        },
        "ece": {
          "value": 0.00826415120785431,
          "lo": 0.006670393035826101,
          "hi": 0.014237934086454946
        },
        "reliability": {
          "value": 0.00015128253894199073,
          "lo": 7.579447670100361e-05,
          "hi": 0.0004981111013192051
        },
        "resolution": {
          "value": 0.0067872674046922975,
          "lo": 0.005359476950057658,
          "hi": 0.008495357238412838
        },
        "uncertainty": {
          "value": 0.0516360003226966,
          "lo": 0.04713235430009503,
          "hi": 0.056413780293513655
        },
        "auroc": {
          "value": 0.8366656995316722,
          "lo": 0.8175586056341942,
          "hi": 0.8558524171817646
        },
        "acc@80": {
          "value": 0.9781639243426111,
          "lo": 0.9741500219637983,
          "hi": 0.9818828417002816
        }
      }
    },
    {
      "env": "boolq",
      "slice": "all",
      "system": "laya-0.1.6 +T",
      "metrics": {
        "n": {
          "value": 8129.0,
          "lo": 7984.475,
          "hi": 8270.575
        },
        "ce": {
          "value": 0.16842074844908808,
          "lo": 0.15377416039538502,
          "hi": 0.1826377502844351
        },
        "kl": {
          "value": 0.16842074844908808,
          "lo": 0.15377416039538502,
          "hi": 0.1826377502844351
        },
        "jsd": {
          "value": 0.039848000608692,
          "lo": 0.03717194335215928,
          "hi": 0.04255754852113439
        },
        "brier": {
          "value": 0.08911969369012436,
          "lo": 0.0807434104771179,
          "hi": 0.09722725633968095
        },
        "acc": {
          "value": 0.9453807356378399,
          "lo": 0.9399843400447427,
          "hi": 0.9504083064841119
        },
        "conf": {
          "value": 0.9547045580034629,
          "lo": 0.9529434187871263,
          "hi": 0.9564472385719919
        },
        "ece": {
          "value": 0.010118889872109627,
          "lo": 0.008057267182127393,
          "hi": 0.016097098115100707
        },
        "reliability": {
          "value": 0.00029446399075773173,
          "lo": 0.00014625249070317913,
          "hi": 0.0007716198702540654
        },
        "resolution": {
          "value": 0.006787507272524246,
          "lo": 0.0053595468591614666,
          "hi": 0.008492841986105507
        },
        "uncertainty": {
          "value": 0.0516360003226966,
          "lo": 0.04713235430009503,
          "hi": 0.056413780293513655
        },
        "auroc": {
          "value": 0.836659545036253,
          "lo": 0.8175513786555788,
          "hi": 0.8558356541920423
        },
        "acc@80": {
          "value": 0.9781639243426111,
          "lo": 0.9741500219637983,
          "hi": 0.9818828417002816
        }
      }
    },
    {
      "env": "chaosnli",
      "slice": "all",
      "system": "base rate",
      "metrics": {
        "n": {
          "value": 2763.0,
          "lo": 2732.475,
          "hi": 2798.525
        },
        "ce": {
          "value": 0.9379373340497879,
          "lo": 0.9285802317593829,
          "hi": 0.9475822940904634
        },
        "kl": {
          "value": 0.40757726398052585,
          "lo": 0.39747565842988464,
          "hi": 0.41904554967110874
        },
        "jsd": {
          "value": 0.11304809900072296,
          "lo": 0.11032258933459423,
          "hi": 0.11615128759372625
        },
        "brier": {
          "value": 0.26981895182793947,
          "lo": 0.2626011497084805,
          "hi": 0.2773451971013373
        },
        "acc": {
          "value": 0.48968512486427795,
          "lo": 0.47177673148510973,
          "hi": 0.5089376292721363
        },
        "conf": {
          "value": 0.4905093503234308,
          "lo": 0.4888078368966116,
          "hi": 0.4923186354261798
        },
        "ece": {
          "value": 0.022808835759865427,
          "lo": 0.008771611429390938,
          "hi": 0.04095054029804827
        },
        "reliability": {
          "value": 0.0005754631118155022,
          "lo": 0.00011636701405239725,
          "hi": 0.0018846778177793791
        },
        "resolution": {
          "value": 0.0003291497616916151,
          "lo": 2.4129556481931592e-05,
          "hi": 0.0016044554135521544
        },
        "uncertainty": {
          "value": 0.24989360335093447,
          "lo": 0.24920342118429026,
          "hi": 0.24999970359945467
        },
        "auroc": {
          "value": 0.5195357833655706,
          "lo": 0.4991706758215456,
          "hi": 0.5404608926184091
        },
        "acc@80": {
          "value": 0.49366515837104075,
          "lo": 0.4756861233826964,
          "hi": 0.5150407931169876
        }
      }
    },
    {
      "env": "chaosnli",
      "slice": "all",
      "system": "laya-0.1.6",
      "metrics": {
        "n": {
          "value": 2763.0,
          "lo": 2732.475,
          "hi": 2798.525
        },
        "ce": {
          "value": 1.0285447436275288,
          "lo": 1.0033821554904991,
          "hi": 1.0545519423823069
        },
        "kl": {
          "value": 0.498184673558267,
          "lo": 0.47789097962855864,
          "hi": 0.51921965502745
        },
        "jsd": {
          "value": 0.10085241061940411,
          "lo": 0.09746306336749984,
          "hi": 0.10453247078825088
        },
        "brier": {
          "value": 0.2719412201157911,
          "lo": 0.26132247894674693,
          "hi": 0.2831178041460505
        },
        "acc": {
          "value": 0.6568946796959826,
          "lo": 0.6389596974115177,
          "hi": 0.6737617308149204
        },
        "conf": {
          "value": 0.7850386756747001,
          "lo": 0.7782867783386168,
          "hi": 0.7904756883129792
        },
        "ece": {
          "value": 0.12836848797463352,
          "lo": 0.11764994557569582,
          "hi": 0.1486388444337131
        },
        "reliability": {
          "value": 0.02160372386829519,
          "lo": 0.018122040872542048,
          "hi": 0.02785991771997649
        },
        "resolution": {
          "value": 0.011270827854638914,
          "lo": 0.009226852780802366,
          "hi": 0.016455992401153814
        },
        "uncertainty": {
          "value": 0.22538405948309503,
          "lo": 0.2198068206209398,
          "hi": 0.23069019528126342
        },
        "auroc": {
          "value": 0.62926793830131,
          "lo": 0.6093513301105226,
          "hi": 0.6516518175130473
        },
        "acc@80": {
          "value": 0.6909502262443439,
          "lo": 0.6722383500218386,
          "hi": 0.7085911083203138
        }
      }
    },
    {
      "env": "chaosnli",
      "slice": "all",
      "system": "laya-0.1.6 +T",
      "metrics": {
        "n": {
          "value": 2763.0,
          "lo": 2732.475,
          "hi": 2798.525
        },
        "ce": {
          "value": 0.8251350047631987,
          "lo": 0.8152288405803177,
          "hi": 0.835546065711284
        },
        "kl": {
          "value": 0.29477493469393684,
          "lo": 0.2862985778747161,
          "hi": 0.3033341949032095
        },
        "jsd": {
          "value": 0.08114470031563259,
          "lo": 0.07900899539493031,
          "hi": 0.08340997116660731
        },
        "brier": {
          "value": 0.19441446983318122,
          "lo": 0.18809944062540326,
          "hi": 0.20078332541152275
        },
        "acc": {
          "value": 0.6568946796959826,
          "lo": 0.6389596974115177,
          "hi": 0.6737617308149204
        },
        "conf": {
          "value": 0.5986317700597011,
          "lo": 0.595133404666914,
          "hi": 0.6016171878298637
        },
        "ece": {
          "value": 0.0627802665034211,
          "lo": 0.05626098050792744,
          "hi": 0.08661036301274243
        },
        "reliability": {
          "value": 0.005811743129342419,
          "lo": 0.004608585782396878,
          "hi": 0.010239545437114193
        },
        "resolution": {
          "value": 0.01383844026256437,
          "lo": 0.01116274735607004,
          "hi": 0.019402317467878596
        },
        "uncertainty": {
          "value": 0.22538405948309503,
          "lo": 0.2198068206209398,
          "hi": 0.23069019528126342
        },
        "auroc": {
          "value": 0.6400919436017249,
          "lo": 0.6191042100717233,
          "hi": 0.6627523607997489
        },
        "acc@80": {
          "value": 0.702262443438914,
          "lo": 0.683554107513062,
          "hi": 0.7210812093966812
        }
      }
    },
    {
      "env": "civilcomments",
      "slice": "all",
      "system": "base rate",
      "metrics": {
        "n": {
          "value": 10581.0,
          "lo": 10261.425,
          "hi": 10904.525
        },
        "ce": {
          "value": 0.4254538278813023,
          "lo": 0.4167216532262873,
          "hi": 0.43439171663848053
        },
        "kl": {
          "value": 0.1939710191119612,
          "lo": 0.1898765552212037,
          "hi": 0.19867000295695353
        },
        "jsd": {
          "value": 0.05293960990922591,
          "lo": 0.05205862775499703,
          "hi": 0.05401522441712282
        },
        "brier": {
          "value": 0.10156534944654734,
          "lo": 0.09792364241321323,
          "hi": 0.10541567724997143
        },
        "acc": {
          "value": 0.9116340610528305,
          "lo": 0.906076390941178,
          "hi": 0.9167691033774958
        },
        "conf": {
          "value": 0.8593235781658778,
          "lo": 0.8593235781658777,
          "hi": 0.859323578165878
        },
        "ece": {
          "value": 0.052310482886952725,
          "lo": 0.046752812775300176,
          "hi": 0.05744552521161804
        },
        "reliability": {
          "value": 0.002736386619866174,
          "lo": 0.0021858255137981748,
          "hi": 0.0032999884123147837
        },
        "resolution": {
          "value": 0.0,
          "lo": 0.0,
          "hi": 0.0
        },
        "uncertainty": {
          "value": 0.0805573997811546,
          "lo": 0.07630351442444205,
          "hi": 0.08510196470879164
        },
        "auroc": {
          "value": 0.5,
          "lo": 0.5,
          "hi": 0.5
        },
        "acc@80": {
          "value": 0.9141169521559362,
          "lo": 0.9055380777445347,
          "hi": 0.9173038439837854
        }
      }
    },
    {
      "env": "civilcomments",
      "slice": "all",
      "system": "laya-0.1.6",
      "metrics": {
        "n": {
          "value": 10581.0,
          "lo": 10261.425,
          "hi": 10904.525
        },
        "ce": {
          "value": 0.3810125971120588,
          "lo": 0.37484793949536976,
          "hi": 0.3860455755495298
        },
        "kl": {
          "value": 0.14952978834271774,
          "lo": 0.14694071031932254,
          "hi": 0.15214002028455306
        },
        "jsd": {
          "value": 0.048026748840901134,
          "lo": 0.04717718855909775,
          "hi": 0.04884049072928548
        },
        "brier": {
          "value": 0.07264455665626256,
          "lo": 0.07091948864137936,
          "hi": 0.07446755140854741
        },
        "acc": {
          "value": 0.9363954257631604,
          "lo": 0.9312894156481802,
          "hi": 0.9412408224641303
        },
        "conf": {
          "value": 0.7462928145292993,
          "lo": 0.7432110057199227,
          "hi": 0.7495198525874994
        },
        "ece": {
          "value": 0.19010261123386113,
          "lo": 0.18572869743920553,
          "hi": 0.19434341943002456
        },
        "reliability": {
          "value": 0.042625286127443054,
          "lo": 0.04090385585377438,
          "hi": 0.044403189351862014
        },
        "resolution": {
          "value": 0.013717049821998603,
          "lo": 0.011881205041954722,
          "hi": 0.01598063638905128
        },
        "uncertainty": {
          "value": 0.059559032372989984,
          "lo": 0.05530653638439067,
          "hi": 0.06398943971092431
        },
        "auroc": {
          "value": 0.9171025739927692,
          "lo": 0.9087929385965824,
          "hi": 0.9240171107172537
        },
        "acc@80": {
          "value": 0.9878322504430006,
          "lo": 0.9847184534863704,
          "hi": 0.9904090341015473
        }
      }
    },
    {
      "env": "civilcomments",
      "slice": "all",
      "system": "laya-0.1.6 +T",
      "metrics": {
        "n": {
          "value": 10581.0,
          "lo": 10261.425,
          "hi": 10904.525
        },
        "ce": {
          "value": 0.31924507592150614,
          "lo": 0.3115205115811762,
          "hi": 0.32654836174716545
        },
        "kl": {
          "value": 0.08776226715216504,
          "lo": 0.08515573370008006,
          "hi": 0.09058161011803512
        },
        "jsd": {
          "value": 0.024394595342128112,
          "lo": 0.023688361737283142,
          "hi": 0.025138113681980403
        },
        "brier": {
          "value": 0.04172181358871891,
          "lo": 0.040081896395935374,
          "hi": 0.04352301107833134
        },
        "acc": {
          "value": 0.9363954257631604,
          "lo": 0.9312894156481802,
          "hi": 0.9412408224641303
        },
        "conf": {
          "value": 0.8608904153831163,
          "lo": 0.8574604423635073,
          "hi": 0.8642629707355846
        },
        "ece": {
          "value": 0.07550501038004401,
          "lo": 0.0718289183674466,
          "hi": 0.07971665523074521
        },
        "reliability": {
          "value": 0.009249599786670091,
          "lo": 0.00825301093897232,
          "hi": 0.010361731443868353
        },
        "resolution": {
          "value": 0.013717049821998603,
          "lo": 0.011852341507289721,
          "hi": 0.01598063638905128
        },
        "uncertainty": {
          "value": 0.059559032372989984,
          "lo": 0.05530653638439067,
          "hi": 0.06398943971092431
        },
        "auroc": {
          "value": 0.9170985248536161,
          "lo": 0.9087881215216295,
          "hi": 0.9240207230291888
        },
        "acc@80": {
          "value": 0.9878322504430006,
          "lo": 0.9847184534863704,
          "hi": 0.9904090341015473
        }
      }
    },
    {
      "env": "folktables",
      "slice": "all",
      "system": "base rate",
      "metrics": {
        "n": {
          "value": 15003.0,
          "lo": 14855.8,
          "hi": 15157.0
        },
        "ce": {
          "value": 0.6832870834397348,
          "lo": 0.6811854548087254,
          "hi": 0.6856621653799928
        },
        "kl": {
          "value": 0.6832870834397348,
          "lo": 0.6811854548087254,
          "hi": 0.6856621653799928
        },
        "jsd": {
          "value": 0.21218818553265442,
          "lo": 0.211602993626582,
          "hi": 0.21284537334357836
        },
        "brier": {
          "value": 0.49029879579639624,
          "lo": 0.4882531508828007,
          "hi": 0.4926056356460426
        },
        "acc": {
          "value": 0.6536026128107711,
          "lo": 0.6451184668229782,
          "hi": 0.661609588316768
        },
        "conf": {
          "value": 0.5285609544757716,
          "lo": 0.5272976034555974,
          "hi": 0.5298136221174654
        },
        "ece": {
          "value": 0.12504165833499967,
          "lo": 0.11694450254962907,
          "hi": 0.13293473708420933
        },
        "reliability": {
          "value": 0.018745024303201244,
          "lo": 0.016591111933824162,
          "hi": 0.02095019662130334
        },
        "resolution": {
          "value": 1.863742707408628e-06,
          "lo": 1.2375750882431937e-08,
          "hi": 9.316471375822222e-05
        },
        "uncertainty": {
          "value": 0.22640623733770432,
          "lo": 0.22388234090058073,
          "hi": 0.2289406268460248
        },
        "auroc": {
          "value": 0.5012058742372862,
          "lo": 0.49426236500946596,
          "hi": 0.5081097296397117
        },
        "acc@80": {
          "value": 0.6346442259623396,
          "lo": 0.64362479180161,
          "hi": 0.6630202666626369
        }
      }
    },
    {
      "env": "folktables",
      "slice": "all",
      "system": "laya-0.1.6",
      "metrics": {
        "n": {
          "value": 15003.0,
          "lo": 14855.8,
          "hi": 15157.0
        },
        "ce": {
          "value": 0.720976727971998,
          "lo": 0.7187275568573948,
          "hi": 0.7231376246748215
        },
        "kl": {
          "value": 0.720976727971998,
          "lo": 0.7187275568573948,
          "hi": 0.7231376246748215
        },
        "jsd": {
          "value": 0.22245330451912865,
          "lo": 0.22185146115893456,
          "hi": 0.22303308672417152
        },
        "brier": {
          "value": 0.5273739882830246,
          "lo": 0.5251673418866659,
          "hi": 0.5294821933633662
        },
        "acc": {
          "value": 0.4391121775644871,
          "lo": 0.4309876348072268,
          "hi": 0.4470842753273963
        },
        "conf": {
          "value": 0.5516842689218898,
          "lo": 0.5510766443317116,
          "hi": 0.5522637898698728
        },
        "ece": {
          "value": 0.11439187478790329,
          "lo": 0.10666349685369207,
          "hi": 0.12235665432272713
        },
        "reliability": {
          "value": 0.019709030457048863,
          "lo": 0.017700443867459795,
          "hi": 0.022151985907463968
        },
        "resolution": {
          "value": 0.002366054834422019,
          "lo": 0.001883000473039606,
          "hi": 0.003348470989642238
        },
        "uncertainty": {
          "value": 0.24629267307906147,
          "lo": 0.24523729330617874,
          "hi": 0.24719992563234872
        },
        "auroc": {
          "value": 0.44683860462548985,
          "lo": 0.43852322856093934,
          "hi": 0.4566225509724676
        },
        "acc@80": {
          "value": 0.4237627062156307,
          "lo": 0.4155537377531345,
          "hi": 0.43319309264175665
        }
      }
    },
    {
      "env": "folktables",
      "slice": "all",
      "system": "laya-0.1.6 +T",
      "metrics": {
        "n": {
          "value": 15003.0,
          "lo": 14855.8,
          "hi": 15157.0
        },
        "ce": {
          "value": 0.6932795192425513,
          "lo": 0.6932643598725224,
          "hi": 0.6932936000889979
        },
        "kl": {
          "value": 0.6932795192425513,
          "lo": 0.6932643598725224,
          "hi": 0.6932936000889979
        },
        "jsd": {
          "value": 0.21579786082603852,
          "lo": 0.21579369688263556,
          "hi": 0.21580172783369722
        },
        "brier": {
          "value": 0.500132338595746,
          "lo": 0.5001171792366494,
          "hi": 0.5001464194336274
        },
        "acc": {
          "value": 0.4391121775644871,
          "lo": 0.4309876348072268,
          "hi": 0.4470842753273963
        },
        "conf": {
          "value": 0.5003518570537647,
          "lo": 0.5003476356222193,
          "hi": 0.500355839523348
        },
        "ece": {
          "value": 0.06343668058765774,
          "lo": 0.05648981460136645,
          "hi": 0.0717494721729754
        },
        "reliability": {
          "value": 0.00613941541133633,
          "lo": 0.005200797478803559,
          "hi": 0.007625049705925511
        },
        "resolution": {
          "value": 0.0023660704975580994,
          "lo": 0.0018832496360224036,
          "hi": 0.0033462913468053673
        },
        "uncertainty": {
          "value": 0.24629267307906147,
          "lo": 0.24523729330617874,
          "hi": 0.24719992563234872
        },
        "auroc": {
          "value": 0.44683114584539635,
          "lo": 0.4385128537020856,
          "hi": 0.45661674335469843
        },
        "acc@80": {
          "value": 0.42367938676887185,
          "lo": 0.4155120588620882,
          "hi": 0.43323697795480515
        }
      }
    },
    {
      "env": "mcq",
      "slice": "all",
      "system": "base rate",
      "metrics": {
        "n": {
          "value": 5064.0,
          "lo": 5064.0,
          "hi": 5064.0
        },
        "ce": {
          "value": 1.4388183798049623,
          "lo": 1.4361334266007217,
          "hi": 1.4412926929165724
        },
        "kl": {
          "value": 1.4388183798049623,
          "lo": 1.4361334266007217,
          "hi": 1.4412926929165724
        },
        "jsd": {
          "value": 0.39036904499300623,
          "lo": 0.3898617116314181,
          "hi": 0.39084013172225024
        },
        "brier": {
          "value": 0.7617430226434967,
          "lo": 0.7611488776987889,
          "hi": 0.7623009643233282
        },
        "acc": {
          "value": 0.2401263823064771,
          "lo": 0.228870458135861,
          "hi": 0.2513921800947867
        },
        "conf": {
          "value": 0.23825697735650342,
          "lo": 0.23769903567667192,
          "hi": 0.23885112230121117
        },
        "ece": {
          "value": 0.0018694049499736714,
          "lo": 0.0012241640337019553,
          "hi": 0.014866870721432324
        },
        "reliability": {
          "value": 8.073839816032738e-05,
          "lo": 1.2232715420221261e-05,
          "hi": 0.0005723834895640047
        },
        "resolution": {
          "value": 0.0006050685727678989,
          "lo": 0.0001906606623764816,
          "hi": 0.0014383353490246106
        },
        "uncertainty": {
          "value": 0.1824657028268807,
          "lo": 0.17648877152854212,
          "hi": 0.18819411298413483
        },
        "auroc": {
          "value": 0.5260262080780173,
          "lo": 0.5131244477517856,
          "hi": 0.5401024591295606
        },
        "acc@80": {
          "value": 0.2512959763021476,
          "lo": 0.23684892619106396,
          "hi": 0.26289804986423104
        }
      }
    },
    {
      "env": "mcq",
      "slice": "all",
      "system": "laya-0.1.6",
      "metrics": {
        "n": {
          "value": 5064.0,
          "lo": 5064.0,
          "hi": 5064.0
        },
        "ce": {
          "value": 1.1993251171399195,
          "lo": 1.176848157016128,
          "hi": 1.2253531746697932
        },
        "kl": {
          "value": 1.1993251171399195,
          "lo": 1.176848157016128,
          "hi": 1.2253531746697932
        },
        "jsd": {
          "value": 0.3027824241850154,
          "lo": 0.298267374595891,
          "hi": 0.30751843367633935
        },
        "brier": {
          "value": 0.6312290009042739,
          "lo": 0.6193712176952356,
          "hi": 0.6439454266304352
        },
        "acc": {
          "value": 0.5078988941548184,
          "lo": 0.4945645734597156,
          "hi": 0.5199447077409163
        },
        "conf": {
          "value": 0.5662472816198876,
          "lo": 0.5608019186667127,
          "hi": 0.5720023136665401
        },
        "ece": {
          "value": 0.06579449038075703,
          "lo": 0.058592720388318593,
          "hi": 0.08158590762440474
        },
        "reliability": {
          "value": 0.007559025677366429,
          "lo": 0.0060630002870996486,
          "hi": 0.01051744168424885
        },
        "resolution": {
          "value": 0.022468520685816,
          "lo": 0.019473980735299188,
          "hi": 0.026768319076778915
        },
        "uncertainty": {
          "value": 0.24993760747113097,
          "lo": 0.24960220863312943,
          "hi": 0.24999996100466945
        },
        "auroc": {
          "value": 0.667545944846214,
          "lo": 0.6533370729818624,
          "hi": 0.6822790215851574
        },
        "acc@80": {
          "value": 0.5573932362379659,
          "lo": 0.5412182177240187,
          "hi": 0.5717106887188349
        }
      }
    },
    {
      "env": "mcq",
      "slice": "all",
      "system": "laya-0.1.6 +T",
      "metrics": {
        "n": {
          "value": 5064.0,
          "lo": 5064.0,
          "hi": 5064.0
        },
        "ce": {
          "value": 1.1787822993133397,
          "lo": 1.1625312940057417,
          "hi": 1.1965745610113971
        },
        "kl": {
          "value": 1.1787822993133397,
          "lo": 1.1625312940057417,
          "hi": 1.1965745610113971
        },
        "jsd": {
          "value": 0.31641723619898476,
          "lo": 0.31287769510313346,
          "hi": 0.32002992408234404
        },
        "brier": {
          "value": 0.6249379778249716,
          "lo": 0.6157089328560164,
          "hi": 0.6343573188963025
        },
        "acc": {
          "value": 0.5078988941548184,
          "lo": 0.4945645734597156,
          "hi": 0.5199447077409163
        },
        "conf": {
          "value": 0.47546104819028784,
          "lo": 0.4711100074624925,
          "hi": 0.4799136990918725
        },
        "ece": {
          "value": 0.0418016715069125,
          "lo": 0.034856876299361336,
          "hi": 0.056485500357719466
        },
        "reliability": {
          "value": 0.0024197159470087408,
          "lo": 0.001863533060433718,
          "hi": 0.0045085865560292124
        },
        "resolution": {
          "value": 0.02228669792844079,
          "lo": 0.019275008490941038,
          "hi": 0.026687375477648448
        },
        "uncertainty": {
          "value": 0.24993760747113097,
          "lo": 0.24960220863312943,
          "hi": 0.24999996100466945
        },
        "auroc": {
          "value": 0.6674585735005205,
          "lo": 0.6531285905081086,
          "hi": 0.6821516095991993
        },
        "acc@80": {
          "value": 0.5568995309800049,
          "lo": 0.5412182177240187,
          "hi": 0.5697358676869909
        }
      }
    },
    {
      "env": "stsb",
      "slice": "all",
      "system": "base rate",
      "metrics": {
        "n": {
          "value": 3676.0,
          "lo": 3150.95,
          "hi": 4793.674999999996
        },
        "ce": {
          "value": 1.7269663984515509,
          "lo": 1.705291022662975,
          "hi": 1.7506263143965053
        },
        "kl": {
          "value": 1.3246625755892953,
          "lo": 1.2976795900285005,
          "hi": 1.3546742057651315
        },
        "jsd": {
          "value": 0.35163188140152485,
          "lo": 0.345662174043538,
          "hi": 0.3583488621411241
        },
        "brier": {
          "value": 0.537933427501715,
          "lo": 0.5253465273948369,
          "hi": 0.5509865061000444
        },
        "acc": {
          "value": 0.24183895538628944,
          "lo": 0.2057694485131953,
          "hi": 0.27246244499642425
        },
        "conf": {
          "value": 0.23596666484615023,
          "lo": 0.2314976047091242,
          "hi": 0.23947197861216712
        },
        "ece": {
          "value": 0.022163881119620442,
          "lo": 0.009124136796962445,
          "hi": 0.04684933352937424
        },
        "reliability": {
          "value": 0.0004970487369445307,
          "lo": 0.0001251405423291301,
          "hi": 0.00260283556603496
        },
        "resolution": {
          "value": 0.001650002222171168,
          "lo": 0.0001834599765092166,
          "hi": 0.00383292294125337
        },
        "uncertainty": {
          "value": 0.18335287504395775,
          "lo": 0.16342811435849147,
          "hi": 0.19822650163650427
        },
        "auroc": {
          "value": 0.5479298268556043,
          "lo": 0.515657963419467,
          "hi": 0.5753557516756425
        },
        "acc@80": {
          "value": 0.2471948316899014,
          "lo": 0.20855173667103932,
          "hi": 0.2797579046695503
        }
      }
    },
    {
      "env": "stsb",
      "slice": "all",
      "system": "laya-0.1.6",
      "metrics": {
        "n": {
          "value": 3676.0,
          "lo": 3150.95,
          "hi": 4793.674999999996
        },
        "ce": {
          "value": 1.6736087922378693,
          "lo": 1.6454919880841663,
          "hi": 1.7003488444969814
        },
        "kl": {
          "value": 1.2713049693756135,
          "lo": 1.2456231458092684,
          "hi": 1.3005210492731487
        },
        "jsd": {
          "value": 0.31901269603382404,
          "lo": 0.3090945377346262,
          "hi": 0.32777996558039413
        },
        "brier": {
          "value": 0.5437303126659617,
          "lo": 0.5326542358620336,
          "hi": 0.5563186476397121
        },
        "acc": {
          "value": 0.28917301414581065,
          "lo": 0.2619405832349143,
          "hi": 0.3228819749910248
        },
        "conf": {
          "value": 0.45360173306897106,
          "lo": 0.43005428967076675,
          "hi": 0.486999864881288
        },
        "ece": {
          "value": 0.16442871892316038,
          "lo": 0.15040931743295344,
          "hi": 0.17809551965658568
        },
        "reliability": {
          "value": 0.0333625174345907,
          "lo": 0.02894269456847073,
          "hi": 0.040629612785721775
        },
        "resolution": {
          "value": 0.016751470987512153,
          "lo": 0.012073706086188617,
          "hi": 0.02212108688598159
        },
        "uncertainty": {
          "value": 0.20555198203563743,
          "lo": 0.1933277138691961,
          "hi": 0.21862919761213467
        },
        "auroc": {
          "value": 0.6473404019773771,
          "lo": 0.6229505005582147,
          "hi": 0.6655828886182101
        },
        "acc@80": {
          "value": 0.3138388303298198,
          "lo": 0.2825246786719,
          "hi": 0.35257716166005415
        }
      }
    },
    {
      "env": "stsb",
      "slice": "all",
      "system": "laya-0.1.6 +T",
      "metrics": {
        "n": {
          "value": 3676.0,
          "lo": 3150.95,
          "hi": 4793.674999999996
        },
        "ce": {
          "value": 1.6062735114762483,
          "lo": 1.5746073747409106,
          "hi": 1.6321932664093877
        },
        "kl": {
          "value": 1.2039696886139928,
          "lo": 1.1755094929077654,
          "hi": 1.2319700579142734
        },
        "jsd": {
          "value": 0.3226950954777939,
          "lo": 0.31440668412059836,
          "hi": 0.32976379597199573
        },
        "brier": {
          "value": 0.5022439281949195,
          "lo": 0.49011437005747943,
          "hi": 0.5151003741658788
        },
        "acc": {
          "value": 0.28917301414581065,
          "lo": 0.2619405832349143,
          "hi": 0.3228819749910248
        },
        "conf": {
          "value": 0.3376014139837463,
          "lo": 0.3221646652656687,
          "hi": 0.3595460530741712
        },
        "ece": {
          "value": 0.04980471678034271,
          "lo": 0.04039839900599371,
          "hi": 0.06840069816459508
        },
        "reliability": {
          "value": 0.0032059262837748044,
          "lo": 0.00261795207411194,
          "hi": 0.006214947101106022
        },
        "resolution": {
          "value": 0.0168465680778093,
          "lo": 0.012226649179908476,
          "hi": 0.02242040477662038
        },
        "uncertainty": {
          "value": 0.20555198203563743,
          "lo": 0.1933277138691961,
          "hi": 0.21862919761213467
        },
        "auroc": {
          "value": 0.6481324472506849,
          "lo": 0.6241612697813905,
          "hi": 0.6662312023612728
        },
        "acc@80": {
          "value": 0.31553893233594016,
          "lo": 0.28261010446385143,
          "hi": 0.35400641765209856
        }
      }
    },
    {
      "env": "yelp",
      "slice": "all",
      "system": "base rate",
      "metrics": {
        "n": {
          "value": 5000.0,
          "lo": 5000.0,
          "hi": 5000.0
        },
        "ce": {
          "value": 1.6094379124340998,
          "lo": 1.6094379124340998,
          "hi": 1.6094379124340998
        },
        "kl": {
          "value": 1.6094379124340998,
          "lo": 1.6094379124340998,
          "hi": 1.6094379124340998
        },
        "jsd": {
          "value": 0.4228104552401622,
          "lo": 0.4228104552401622,
          "hi": 0.4228104552401622
        },
        "brier": {
          "value": 0.8000000000000004,
          "lo": 0.8000000000000003,
          "hi": 0.8000000000000004
        },
        "acc": {
          "value": 0.2,
          "lo": 0.18939,
          "hi": 0.211505
        },
        "conf": {
          "value": 0.19999999999999998,
          "lo": 0.19999999999999998,
          "hi": 0.19999999999999998
        },
        "ece": {
          "value": 2.7755575615628914e-17,
          "lo": 0.0003999999999999837,
          "hi": 0.012600000000000028
        },
        "reliability": {
          "value": 7.703719777548943e-34,
          "lo": 1.5999999999998695e-07,
          "hi": 0.0001587600000000007
        },
        "resolution": {
          "value": 0.0,
          "lo": 0.0,
          "hi": 0.0
        },
        "uncertainty": {
          "value": 0.16000000000000003,
          "lo": 0.15352138799999998,
          "hi": 0.166770625
        },
        "auroc": {
          "value": 0.5,
          "lo": 0.5,
          "hi": 0.5
        },
        "acc@80": {
          "value": 0.25,
          "lo": 0.18761875,
          "hi": 0.2127625
        }
      }
    },
    {
      "env": "yelp",
      "slice": "all",
      "system": "laya-0.1.6",
      "metrics": {
        "n": {
          "value": 5000.0,
          "lo": 5000.0,
          "hi": 5000.0
        },
        "ce": {
          "value": 1.1845529684192286,
          "lo": 1.1431186098319637,
          "hi": 1.221813557092446
        },
        "kl": {
          "value": 1.1845529684192286,
          "lo": 1.1431186098319637,
          "hi": 1.221813557092446
        },
        "jsd": {
          "value": 0.2648056976124352,
          "lo": 0.25859518272782595,
          "hi": 0.27073400111055734
        },
        "brier": {
          "value": 0.5980090902700684,
          "lo": 0.5806929323199541,
          "hi": 0.6144723507243347
        },
        "acc": {
          "value": 0.5624,
          "lo": 0.548095,
          "hi": 0.576505
        },
        "conf": {
          "value": 0.6928905481040478,
          "lo": 0.6877329366482795,
          "hi": 0.6983667400699853
        },
        "ece": {
          "value": 0.13049054810404778,
          "lo": 0.1181634066042304,
          "hi": 0.14539070894256234
        },
        "reliability": {
          "value": 0.021255738396660482,
          "lo": 0.01769159733360709,
          "hi": 0.026098285027633436
        },
        "resolution": {
          "value": 0.02356813782596968,
          "lo": 0.020821619069272403,
          "hi": 0.02775570260559374
        },
        "uncertainty": {
          "value": 0.24610624,
          "lo": 0.24414697500000002,
          "hi": 0.247686861
        },
        "auroc": {
          "value": 0.6714477454939786,
          "lo": 0.6578699212962908,
          "hi": 0.6860522557179326
        },
        "acc@80": {
          "value": 0.6115,
          "lo": 0.59625,
          "hi": 0.6278812500000001
        }
      }
    },
    {
      "env": "yelp",
      "slice": "all",
      "system": "laya-0.1.6 +T",
      "metrics": {
        "n": {
          "value": 5000.0,
          "lo": 5000.0,
          "hi": 5000.0
        },
        "ce": {
          "value": 1.0798171229857636,
          "lo": 1.0550082317900924,
          "hi": 1.1030971551084312
        },
        "kl": {
          "value": 1.0798171229857636,
          "lo": 1.0550082317900924,
          "hi": 1.1030971551084312
        },
        "jsd": {
          "value": 0.28105126217439047,
          "lo": 0.27631426430754846,
          "hi": 0.2855167871962946
        },
        "brier": {
          "value": 0.5691607375831819,
          "lo": 0.5563139269221727,
          "hi": 0.5804567829613643
        },
        "acc": {
          "value": 0.5624,
          "lo": 0.548095,
          "hi": 0.576505
        },
        "conf": {
          "value": 0.5649238940779658,
          "lo": 0.5602919850396543,
          "hi": 0.569808322521585
        },
        "ece": {
          "value": 0.023761205308505857,
          "lo": 0.022986379472374863,
          "hi": 0.043442903484244774
        },
        "reliability": {
          "value": 0.0008859507295626262,
          "lo": 0.0008034458816503666,
          "hi": 0.0028251749879963177
        },
        "resolution": {
          "value": 0.0240286258589128,
          "lo": 0.02139084002659621,
          "hi": 0.02833310805383938
        },
        "uncertainty": {
          "value": 0.24610624,
          "lo": 0.24414697500000002,
          "hi": 0.247686861
        },
        "auroc": {
          "value": 0.673008209787773,
          "lo": 0.6592879328838852,
          "hi": 0.6876161819228981
        },
        "acc@80": {
          "value": 0.60925,
          "lo": 0.59436875,
          "hi": 0.6258812500000001
        }
      }
    },
    {
      "env": "boolq",
      "slice": "all",
      "system": "qwen3.8-27b",
      "metrics": {
        "n": {
          "value": 8129.0,
          "lo": 7984.475,
          "hi": 8270.575
        },
        "ce": {
          "value": 0.38917437609910194,
          "lo": 0.36478176193977996,
          "hi": 0.4117113811120201
        },
        "kl": {
          "value": 0.38917437609910194,
          "lo": 0.36478176193977996,
          "hi": 0.4117113811120201
        },
        "jsd": {
          "value": 0.07719183345624069,
          "lo": 0.07339144015470589,
          "hi": 0.08092392101720004
        },
        "brier": {
          "value": 0.20377173136231924,
          "lo": 0.19210978848275312,
          "hi": 0.21488258715437442
        },
        "acc": {
          "value": 0.8737852134333867,
          "lo": 0.8662839492571259,
          "hi": 0.8814999082101197
        },
        "conf": {
          "value": 0.9438369756317792,
          "lo": 0.9414265075098082,
          "hi": 0.9461442786298321
        },
        "ece": {
          "value": 0.07005176219839243,
          "lo": 0.06288961645106496,
          "hi": 0.07734588351195741
        },
        "reliability": {
          "value": 0.007471879521472837,
          "lo": 0.006190243666681738,
          "hi": 0.00926393843997429
        },
        "resolution": {
          "value": 0.015923177972556827,
          "lo": 0.013562333586149574,
          "hi": 0.018614559575419718
        },
        "uncertainty": {
          "value": 0.11028461421855758,
          "lo": 0.10445781851909179,
          "hi": 0.11583606612118956
        },
        "auroc": {
          "value": 0.7923272680269353,
          "lo": 0.7754385655654847,
          "hi": 0.806692445445645
        },
        "acc@80": {
          "value": 0.9298785176072581,
          "lo": 0.9236048483392075,
          "hi": 0.9361573239432275
        }
      }
    },
    {
      "env": "boolq",
      "slice": "all",
      "system": "qwen3.8-27b +T",
      "metrics": {
        "n": {
          "value": 8129.0,
          "lo": 7984.475,
          "hi": 8270.575
        },
        "ce": {
          "value": 0.31589675014113305,
          "lo": 0.3022209294312514,
          "hi": 0.3288799637631647
        },
        "kl": {
          "value": 0.31589675014113305,
          "lo": 0.3022209294312514,
          "hi": 0.3288799637631647
        },
        "jsd": {
          "value": 0.08575448318811,
          "lo": 0.08262998191879009,
          "hi": 0.08875348138449343
        },
        "brier": {
          "value": 0.1893347868894183,
          "lo": 0.1799530596178881,
          "hi": 0.19822644898122355
        },
        "acc": {
          "value": 0.8737852134333867,
          "lo": 0.8662839492571259,
          "hi": 0.8814999082101197
        },
        "conf": {
          "value": 0.8793572729627128,
          "lo": 0.876571291263446,
          "hi": 0.882064924721078
        },
        "ece": {
          "value": 0.015191298573106632,
          "lo": 0.01243495301594864,
          "hi": 0.022962742675946134
        },
        "reliability": {
          "value": 0.0003958249808374941,
          "lo": 0.00024787725473796393,
          "hi": 0.0009551273630832529
        },
        "resolution": {
          "value": 0.015923177972556827,
          "lo": 0.013562333586149574,
          "hi": 0.018614559575419718
        },
        "uncertainty": {
          "value": 0.11028461421855758,
          "lo": 0.10445781851909179,
          "hi": 0.11583606612118956
        },
        "auroc": {
          "value": 0.7923272680269353,
          "lo": 0.7754385655654847,
          "hi": 0.806692445445645
        },
        "acc@80": {
          "value": 0.9298785176072581,
          "lo": 0.9236048483392075,
          "hi": 0.9361573239432275
        }
      }
    },
    {
      "env": "chaosnli",
      "slice": "all",
      "system": "qwen3.8-27b",
      "metrics": {
        "n": {
          "value": 2763.0,
          "lo": 2732.475,
          "hi": 2798.525
        },
        "ce": {
          "value": 1.1717923312186944,
          "lo": 1.138586864694005,
          "hi": 1.20638100834521
        },
        "kl": {
          "value": 0.6414322611494326,
          "lo": 0.6127829502553543,
          "hi": 0.672345413595422
        },
        "jsd": {
          "value": 0.07691632010275923,
          "lo": 0.07386184330292474,
          "hi": 0.0798222751056858
        },
        "brier": {
          "value": 0.18196941143197648,
          "lo": 0.1719848329072861,
          "hi": 0.1914884979845514
        },
        "acc": {
          "value": 0.7604053564965617,
          "lo": 0.7432688129163535,
          "hi": 0.7773603885998843
        },
        "conf": {
          "value": 0.8891980318490847,
          "lo": 0.8833185849628782,
          "hi": 0.8948141919312281
        },
        "ece": {
          "value": 0.12880139700185403,
          "lo": 0.11318601239529825,
          "hi": 0.14434011515062262
        },
        "reliability": {
          "value": 0.022344589657931863,
          "lo": 0.017880110274394165,
          "hi": 0.02871049996449803
        },
        "resolution": {
          "value": 0.03430066836980699,
          "lo": 0.029370157179605578,
          "hi": 0.04077879134663945
        },
        "uncertainty": {
          "value": 0.18218905030789861,
          "lo": 0.17307121157914737,
          "hi": 0.1908196954129349
        },
        "auroc": {
          "value": 0.7887252653390487,
          "lo": 0.7712352266239298,
          "hi": 0.8048739902785704
        },
        "acc@80": {
          "value": 0.8307692307692308,
          "lo": 0.8124315473649547,
          "hi": 0.848554618482514
        }
      }
    },
    {
      "env": "chaosnli",
      "slice": "all",
      "system": "qwen3.8-27b +T",
      "metrics": {
        "n": {
          "value": 2763.0,
          "lo": 2732.475,
          "hi": 2798.525
        },
        "ce": {
          "value": 0.7060486628596031,
          "lo": 0.6928300845033976,
          "hi": 0.7187495942850437
        },
        "kl": {
          "value": 0.1756885927903413,
          "lo": 0.16822572806269168,
          "hi": 0.18281272188519052
        },
        "jsd": {
          "value": 0.04626507343288381,
          "lo": 0.04436686621685745,
          "hi": 0.047964706815352545
        },
        "brier": {
          "value": 0.09934163650724467,
          "lo": 0.09412590074193818,
          "hi": 0.1043263450093202
        },
        "acc": {
          "value": 0.7604053564965617,
          "lo": 0.7432688129163535,
          "hi": 0.7773603885998843
        },
        "conf": {
          "value": 0.6842320865266962,
          "lo": 0.6781952964143029,
          "hi": 0.6904375285745871
        },
        "ece": {
          "value": 0.07858697628390005,
          "lo": 0.06939970982409929,
          "hi": 0.09493126035293394
        },
        "reliability": {
          "value": 0.007777248597400178,
          "lo": 0.0062565383718330165,
          "hi": 0.011154558191824553
        },
        "resolution": {
          "value": 0.03757447751388948,
          "lo": 0.03259328935851462,
          "hi": 0.04455272036935069
        },
        "uncertainty": {
          "value": 0.18218905030789861,
          "lo": 0.17307121157914737,
          "hi": 0.1908196954129349
        },
        "auroc": {
          "value": 0.801060781012063,
          "lo": 0.7834037982866692,
          "hi": 0.8174479544838923
        },
        "acc@80": {
          "value": 0.8380090497737557,
          "lo": 0.819881352229595,
          "hi": 0.8555053007886539
        }
      }
    },
    {
      "env": "civilcomments",
      "slice": "all",
      "system": "qwen3.8-27b",
      "metrics": {
        "n": {
          "value": 10581.0,
          "lo": 10261.425,
          "hi": 10904.525
        },
        "ce": {
          "value": 0.5176786123346133,
          "lo": 0.505453019906735,
          "hi": 0.530276195831063
        },
        "kl": {
          "value": 0.28619580356527224,
          "lo": 0.27642535835701254,
          "hi": 0.29615644767286525
        },
        "jsd": {
          "value": 0.06800278814294652,
          "lo": 0.06598507291823116,
          "hi": 0.07004084242617473
        },
        "brier": {
          "value": 0.17101015392468566,
          "lo": 0.1643534221508666,
          "hi": 0.17728369471419955
        },
        "acc": {
          "value": 0.8048388621113316,
          "lo": 0.7953190817461367,
          "hi": 0.8132878036263456
        },
        "conf": {
          "value": 0.841425401683969,
          "lo": 0.8382532675599482,
          "hi": 0.8447274094378437
        },
        "ece": {
          "value": 0.040925964535554055,
          "lo": 0.036076279908205756,
          "hi": 0.04868327493816005
        },
        "reliability": {
          "value": 0.0031079957315440234,
          "lo": 0.002517658730804907,
          "hi": 0.004257526650648568
        },
        "resolution": {
          "value": 0.019630506626700843,
          "lo": 0.01802207771597479,
          "hi": 0.021900322431707055
        },
        "uncertainty": {
          "value": 0.15707326814666855,
          "lo": 0.15185071675929868,
          "hi": 0.16278661948606002
        },
        "auroc": {
          "value": 0.7538875689913418,
          "lo": 0.7437286645947476,
          "hi": 0.763803056702711
        },
        "acc@80": {
          "value": 0.8555227406969876,
          "lo": 0.8464496598759651,
          "hi": 0.8634669478649846
        }
      }
    },
    {
      "env": "civilcomments",
      "slice": "all",
      "system": "qwen3.8-27b +T",
      "metrics": {
        "n": {
          "value": 10581.0,
          "lo": 10261.425,
          "hi": 10904.525
        },
        "ce": {
          "value": 0.47297702112235573,
          "lo": 0.46389071412431704,
          "hi": 0.48134516777486236
        },
        "kl": {
          "value": 0.24149421235301474,
          "lo": 0.2352554985703877,
          "hi": 0.24806821852400143
        },
        "jsd": {
          "value": 0.0695681637074148,
          "lo": 0.06792337761642556,
          "hi": 0.07123778931909612
        },
        "brier": {
          "value": 0.15797512535439767,
          "lo": 0.15300866671717778,
          "hi": 0.16292719893400803
        },
        "acc": {
          "value": 0.8048388621113316,
          "lo": 0.7953190817461367,
          "hi": 0.8132878036263456
        },
        "conf": {
          "value": 0.7659403516461888,
          "lo": 0.7629512527726343,
          "hi": 0.7694311478600653
        },
        "ece": {
          "value": 0.03889851046514287,
          "lo": 0.032308723391677986,
          "hi": 0.04674378130222563
        },
        "reliability": {
          "value": 0.002204465811645611,
          "lo": 0.001631123972190602,
          "hi": 0.003318884561341544
        },
        "resolution": {
          "value": 0.019630506626700843,
          "lo": 0.01802207771597479,
          "hi": 0.021900322431707055
        },
        "uncertainty": {
          "value": 0.15707326814666855,
          "lo": 0.15185071675929868,
          "hi": 0.16278661948606002
        },
        "auroc": {
          "value": 0.7538875974237925,
          "lo": 0.7437286645947476,
          "hi": 0.7638030715603317
        },
        "acc@80": {
          "value": 0.8555227406969876,
          "lo": 0.8464496598759651,
          "hi": 0.8634669478649846
        }
      }
    },
    {
      "env": "folktables",
      "slice": "all",
      "system": "qwen3.8-27b",
      "metrics": {
        "n": {
          "value": 15003.0,
          "lo": 14855.8,
          "hi": 15157.0
        },
        "ce": {
          "value": 0.5152170611479078,
          "lo": 0.5013345977707182,
          "hi": 0.5271323194698895
        },
        "kl": {
          "value": 0.5152170611479078,
          "lo": 0.5013345977707182,
          "hi": 0.5271323194698895
        },
        "jsd": {
          "value": 0.13197355737727537,
          "lo": 0.1292066427756764,
          "hi": 0.13466268637823997
        },
        "brier": {
          "value": 0.3300800164591847,
          "lo": 0.32136915186904,
          "hi": 0.3379782638183155
        },
        "acc": {
          "value": 0.7638472305538893,
          "lo": 0.756615411131695,
          "hi": 0.7708374099107584
        },
        "conf": {
          "value": 0.8475244302376784,
          "lo": 0.8447128637395995,
          "hi": 0.8499934222016116
        },
        "ece": {
          "value": 0.0836771996837894,
          "lo": 0.07735320245887115,
          "hi": 0.09023618321654706
        },
        "reliability": {
          "value": 0.008968703120597428,
          "lo": 0.007738443309191269,
          "hi": 0.010453365822199101
        },
        "resolution": {
          "value": 0.024364708249844,
          "lo": 0.02252781376785893,
          "hi": 0.026466105696258707
        },
        "uncertainty": {
          "value": 0.1803846389290428,
          "lo": 0.17664708837189344,
          "hi": 0.18414851315321487
        },
        "auroc": {
          "value": 0.7486387015864431,
          "lo": 0.7404284055591798,
          "hi": 0.7573377305552835
        },
        "acc@80": {
          "value": 0.8189468421929679,
          "lo": 0.8115530142839764,
          "hi": 0.8258698916099144
        }
      }
    },
    {
      "env": "folktables",
      "slice": "all",
      "system": "qwen3.8-27b +T",
      "metrics": {
        "n": {
          "value": 15003.0,
          "lo": 14855.8,
          "hi": 15157.0
        },
        "ce": {
          "value": 0.4718461850593996,
          "lo": 0.46285303778129794,
          "hi": 0.47994812655713986
        },
        "kl": {
          "value": 0.4718461850593996,
          "lo": 0.46285303778129794,
          "hi": 0.47994812655713986
        },
        "jsd": {
          "value": 0.13745501746238115,
          "lo": 0.13529276541690255,
          "hi": 0.13957112105282737
        },
        "brier": {
          "value": 0.3131132197819545,
          "lo": 0.3064709953074027,
          "hi": 0.3195434958143069
        },
        "acc": {
          "value": 0.7638472305538893,
          "lo": 0.756615411131695,
          "hi": 0.7708374099107584
        },
        "conf": {
          "value": 0.7789741979156999,
          "lo": 0.7761000959587006,
          "hi": 0.7813742400176061
        },
        "ece": {
          "value": 0.01823993969565023,
          "lo": 0.01435277097834691,
          "hi": 0.025505661267580974
        },
        "reliability": {
          "value": 0.0005702561447984441,
          "lo": 0.00037617808288516083,
          "hi": 0.0011406738456206653
        },
        "resolution": {
          "value": 0.024364708249844,
          "lo": 0.02252781376785893,
          "hi": 0.026466105696258707
        },
        "uncertainty": {
          "value": 0.1803846389290428,
          "lo": 0.17664708837189344,
          "hi": 0.18414851315321487
        },
        "auroc": {
          "value": 0.7486387139008708,
          "lo": 0.7404284055591798,
          "hi": 0.7573377305552835
        },
        "acc@80": {
          "value": 0.8189468421929679,
          "lo": 0.8115530142839764,
          "hi": 0.8258698916099144
        }
      }
    },
    {
      "env": "mcq",
      "slice": "all",
      "system": "qwen3.8-27b",
      "metrics": {
        "n": {
          "value": 5064.0,
          "lo": 5064.0,
          "hi": 5064.0
        },
        "ce": {
          "value": 0.16152235565574208,
          "lo": 0.1436146447662271,
          "hi": 0.18117535842745647
        },
        "kl": {
          "value": 0.16152235565574208,
          "lo": 0.1436146447662271,
          "hi": 0.18117535842745647
        },
        "jsd": {
          "value": 0.0339809006728758,
          "lo": 0.03120830008883931,
          "hi": 0.03714233939422152
        },
        "brier": {
          "value": 0.07759509006027641,
          "lo": 0.07031103397817423,
          "hi": 0.08625582464906033
        },
        "acc": {
          "value": 0.9506319115323855,
          "lo": 0.9444065955766193,
          "hi": 0.9558698657187993
        },
        "conf": {
          "value": 0.9612343825055764,
          "lo": 0.9584609354556813,
          "hi": 0.964068326400297
        },
        "ece": {
          "value": 0.011042265252385032,
          "lo": 0.008490387787188679,
          "hi": 0.01711441036696952
        },
        "reliability": {
          "value": 0.00033699148728081663,
          "lo": 0.00018894192159061082,
          "hi": 0.0009361550731056858
        },
        "resolution": {
          "value": 0.010201227058651932,
          "lo": 0.007894381945680142,
          "hi": 0.01323264157178327
        },
        "uncertainty": {
          "value": 0.04693088030866832,
          "lo": 0.04218265580506332,
          "hi": 0.05250276808353859
        },
        "auroc": {
          "value": 0.9187461570419609,
          "lo": 0.9038582629555388,
          "hi": 0.9315262910876058
        },
        "acc@80": {
          "value": 0.9918538632436436,
          "lo": 0.9892557393236238,
          "hi": 0.994569242162429
        }
      }
    },
    {
      "env": "mcq",
      "slice": "all",
      "system": "qwen3.8-27b +T",
      "metrics": {
        "n": {
          "value": 5064.0,
          "lo": 5064.0,
          "hi": 5064.0
        },
        "ce": {
          "value": 0.15856058240573354,
          "lo": 0.14420100372376268,
          "hi": 0.17388267441183627
        },
        "kl": {
          "value": 0.15856058240573354,
          "lo": 0.14420100372376268,
          "hi": 0.17388267441183627
        },
        "jsd": {
          "value": 0.040311464109205934,
          "lo": 0.03765378380738312,
          "hi": 0.043271018983395255
        },
        "brier": {
          "value": 0.07782049892680105,
          "lo": 0.07090973446475114,
          "hi": 0.08542435214527039
        },
        "acc": {
          "value": 0.9506319115323855,
          "lo": 0.9444065955766193,
          "hi": 0.9558698657187993
        },
        "conf": {
          "value": 0.9368245108445022,
          "lo": 0.9332560573742827,
          "hi": 0.9402748198546733
        },
        "ece": {
          "value": 0.014467744564029262,
          "lo": 0.01030803101457266,
          "hi": 0.02049353649834128
        },
        "reliability": {
          "value": 0.0006845060520818079,
          "lo": 0.00031037112769545055,
          "hi": 0.0015105244764829953
        },
        "resolution": {
          "value": 0.010048250389244446,
          "lo": 0.007674740410042408,
          "hi": 0.012867026968922805
        },
        "uncertainty": {
          "value": 0.04693088030866832,
          "lo": 0.04218265580506332,
          "hi": 0.05250276808353859
        },
        "auroc": {
          "value": 0.9181204819277109,
          "lo": 0.9034286574202369,
          "hi": 0.9307128031294514
        },
        "acc@80": {
          "value": 0.992100715872624,
          "lo": 0.9893853369538386,
          "hi": 0.9948160947914095
        }
      }
    },
    {
      "env": "stsb",
      "slice": "all",
      "system": "qwen3.8-27b",
      "metrics": {
        "n": {
          "value": 3676.0,
          "lo": 3150.95,
          "hi": 4793.674999999996
        },
        "ce": {
          "value": 1.6476253216148637,
          "lo": 1.5775778269475285,
          "hi": 1.7187834538113418
        },
        "kl": {
          "value": 1.2453214987526078,
          "lo": 1.1781597293366053,
          "hi": 1.3132812315415179
        },
        "jsd": {
          "value": 0.2531612317538707,
          "lo": 0.23836222277454266,
          "hi": 0.2660907919098357
        },
        "brier": {
          "value": 0.49555985949446896,
          "lo": 0.4665357822966389,
          "hi": 0.5207911062026427
        },
        "acc": {
          "value": 0.43307943416757344,
          "lo": 0.4084860412975492,
          "hi": 0.4593722714036901
        },
        "conf": {
          "value": 0.6409933912226389,
          "lo": 0.6288105489244771,
          "hi": 0.656202620137405
        },
        "ece": {
          "value": 0.20791395705506535,
          "lo": 0.1920427844589958,
          "hi": 0.22585973683509933
        },
        "reliability": {
          "value": 0.05260596751028478,
          "lo": 0.04485848169615247,
          "hi": 0.0640769550047139
        },
        "resolution": {
          "value": 0.0210368610762396,
          "lo": 0.01579295412315698,
          "hi": 0.028579310257640548
        },
        "uncertainty": {
          "value": 0.24552163786866787,
          "lo": 0.2416251851428358,
          "hi": 0.24834907728376013
        },
        "auroc": {
          "value": 0.6278471291196868,
          "lo": 0.5916084195966954,
          "hi": 0.6613640926526575
        },
        "acc@80": {
          "value": 0.458347500850051,
          "lo": 0.4284039479890641,
          "hi": 0.49020417161416496
        }
      }
    },
    {
      "env": "stsb",
      "slice": "all",
      "system": "qwen3.8-27b +T",
      "metrics": {
        "n": {
          "value": 3676.0,
          "lo": 3150.95,
          "hi": 4793.674999999996
        },
        "ce": {
          "value": 1.346727264947815,
          "lo": 1.2981053078549638,
          "hi": 1.387067630833167
        },
        "kl": {
          "value": 0.9444234420855593,
          "lo": 0.9002371630659755,
          "hi": 0.9827978015240808
        },
        "jsd": {
          "value": 0.2544173722121313,
          "lo": 0.2431917264063818,
          "hi": 0.2638236801564275
        },
        "brier": {
          "value": 0.4137424475287545,
          "lo": 0.3945117940173643,
          "hi": 0.43107894651028067
        },
        "acc": {
          "value": 0.43307943416757344,
          "lo": 0.4084860412975492,
          "hi": 0.4593722714036901
        },
        "conf": {
          "value": 0.43821688614180715,
          "lo": 0.4277928172327064,
          "hi": 0.45204815469211607
        },
        "ece": {
          "value": 0.04762515082161155,
          "lo": 0.041572815509467266,
          "hi": 0.0712501592630694
        },
        "reliability": {
          "value": 0.0034786243859689096,
          "lo": 0.002723084325544467,
          "hi": 0.006703017478029771
        },
        "resolution": {
          "value": 0.02259646870336115,
          "lo": 0.017410538777569182,
          "hi": 0.03078626929564409
        },
        "uncertainty": {
          "value": 0.24552163786866787,
          "lo": 0.2416251851428358,
          "hi": 0.24834907728376013
        },
        "auroc": {
          "value": 0.6288149601172851,
          "lo": 0.5940100615585853,
          "hi": 0.6597205074364987
        },
        "acc@80": {
          "value": 0.4546072764365862,
          "lo": 0.42664841306830836,
          "hi": 0.4858779236978728
        }
      }
    },
    {
      "env": "yelp",
      "slice": "all",
      "system": "qwen3.8-27b",
      "metrics": {
        "n": {
          "value": 5000.0,
          "lo": 5000.0,
          "hi": 5000.0
        },
        "ce": {
          "value": 1.0750614394949205,
          "lo": 1.0249181485562968,
          "hi": 1.1214973037621498
        },
        "kl": {
          "value": 1.0750614394949205,
          "lo": 1.0249181485562968,
          "hi": 1.1214973037621498
        },
        "jsd": {
          "value": 0.1992499097762599,
          "lo": 0.1921623077584074,
          "hi": 0.20578540527137232
        },
        "brier": {
          "value": 0.5048983976118063,
          "lo": 0.48400116815630245,
          "hi": 0.52369317141654
        },
        "acc": {
          "value": 0.6704,
          "lo": 0.658095,
          "hi": 0.684305
        },
        "conf": {
          "value": 0.8447078410241782,
          "lo": 0.8402829839710233,
          "hi": 0.8494419245794594
        },
        "ece": {
          "value": 0.17707699045033023,
          "lo": 0.1655175251053024,
          "hi": 0.18917956895743576
        },
        "reliability": {
          "value": 0.040149556280830494,
          "lo": 0.035376448880104855,
          "hi": 0.04608360773679248
        },
        "resolution": {
          "value": 0.02467955631511751,
          "lo": 0.021987147303851054,
          "hi": 0.028814119240178446
        },
        "uncertainty": {
          "value": 0.22096384,
          "lo": 0.216031657,
          "hi": 0.22500596099999998
        },
        "auroc": {
          "value": 0.6987530629445976,
          "lo": 0.6840858584822799,
          "hi": 0.7139086803877877
        },
        "acc@80": {
          "value": 0.71575,
          "lo": 0.7025,
          "hi": 0.7306312500000001
        }
      }
    },
    {
      "env": "yelp",
      "slice": "all",
      "system": "qwen3.8-27b +T",
      "metrics": {
        "n": {
          "value": 5000.0,
          "lo": 5000.0,
          "hi": 5000.0
        },
        "ce": {
          "value": 0.8579646852100378,
          "lo": 0.8307245312991125,
          "hi": 0.8845571792743167
        },
        "kl": {
          "value": 0.8579646852100378,
          "lo": 0.8307245312991125,
          "hi": 0.8845571792743167
        },
        "jsd": {
          "value": 0.21839012188284068,
          "lo": 0.21317031798038433,
          "hi": 0.22358432096686606
        },
        "brier": {
          "value": 0.4560912008380844,
          "lo": 0.4403772714428241,
          "hi": 0.4699115792262956
        },
        "acc": {
          "value": 0.6704,
          "lo": 0.658095,
          "hi": 0.684305
        },
        "conf": {
          "value": 0.6933824652455798,
          "lo": 0.6882470527288771,
          "hi": 0.6985853923447576
        },
        "ece": {
          "value": 0.062183613836976986,
          "lo": 0.05147542975040488,
          "hi": 0.0739244817342388
        },
        "reliability": {
          "value": 0.005187854546029306,
          "lo": 0.0040598069598425345,
          "hi": 0.007699148668794739
        },
        "resolution": {
          "value": 0.023317555991440544,
          "lo": 0.020700572010863134,
          "hi": 0.02715480823596157
        },
        "uncertainty": {
          "value": 0.22096384,
          "lo": 0.216031657,
          "hi": 0.22500596099999998
        },
        "auroc": {
          "value": 0.6935266874435202,
          "lo": 0.6782681124440355,
          "hi": 0.707951431911175
        },
        "acc@80": {
          "value": 0.71275,
          "lo": 0.69911875,
          "hi": 0.72813125
        }
      }
    },
    {
      "env": "boolq",
      "slice": "all",
      "system": "LFM2.5-VL-450M",
      "metrics": {
        "n": {
          "value": 8129.0,
          "lo": 7984.475,
          "hi": 8270.575
        },
        "ce": {
          "value": 1.8442211704383502,
          "lo": 1.7851897229303701,
          "hi": 1.9091822456500516
        },
        "kl": {
          "value": 1.8442211704383502,
          "lo": 1.7851897229303701,
          "hi": 1.9091822456500516
        },
        "jsd": {
          "value": 0.24937887226282604,
          "lo": 0.2419793140327026,
          "hi": 0.2580996156272248
        },
        "brier": {
          "value": 0.7293012228537639,
          "lo": 0.707313065987665,
          "hi": 0.7551299460192672
        },
        "acc": {
          "value": 0.6223397711895682,
          "lo": 0.6091156416180727,
          "hi": 0.6334462215543152
        },
        "conf": {
          "value": 0.9874027616233605,
          "lo": 0.9865438820518876,
          "hi": 0.9880995577569468
        },
        "ece": {
          "value": 0.3650629904337923,
          "lo": 0.35360355814649796,
          "hi": 0.3783027822129629
        },
        "reliability": {
          "value": 0.15031460025611365,
          "lo": 0.1416393213713674,
          "hi": 0.16045640817015988
        },
        "resolution": {
          "value": 0.021194817232827602,
          "lo": 0.018512684938401948,
          "hi": 0.024949449951594642
        },
        "uncertainty": {
          "value": 0.2350329803852841,
          "lo": 0.23219210586793973,
          "hi": 0.23809377671631937
        },
        "auroc": {
          "value": 0.6765497423561583,
          "lo": 0.6622366769117792,
          "hi": 0.6908041518723789
        },
        "acc@80": {
          "value": 0.6730739658619099,
          "lo": 0.6583250142349811,
          "hi": 0.6852436048190568
        }
      }
    },
    {
      "env": "boolq",
      "slice": "all",
      "system": "LFM2.5-VL-450M +T",
      "metrics": {
        "n": {
          "value": 8129.0,
          "lo": 7984.475,
          "hi": 8270.575
        },
        "ce": {
          "value": 0.6462882006413323,
          "lo": 0.6391522517704625,
          "hi": 0.6546312670639313
        },
        "kl": {
          "value": 0.6462882006413323,
          "lo": 0.6391522517704625,
          "hi": 0.6546312670639313
        },
        "jsd": {
          "value": 0.19800455056845404,
          "lo": 0.19609560000277734,
          "hi": 0.2002744331262142
        },
        "brier": {
          "value": 0.4548536484309796,
          "lo": 0.4480256042879124,
          "hi": 0.4628228606147715
        },
        "acc": {
          "value": 0.6223397711895682,
          "lo": 0.6091156416180727,
          "hi": 0.6334462215543152
        },
        "conf": {
          "value": 0.6498392437994798,
          "lo": 0.6489468585458049,
          "hi": 0.6506301063148058
        },
        "ece": {
          "value": 0.09650537872482522,
          "lo": 0.0878912971113201,
          "hi": 0.10872639955944483
        },
        "reliability": {
          "value": 0.013666581300217133,
          "lo": 0.011609939079836616,
          "hi": 0.016744569783581506
        },
        "resolution": {
          "value": 0.021194817232827602,
          "lo": 0.018512684938401948,
          "hi": 0.024949449951594642
        },
        "uncertainty": {
          "value": 0.2350329803852841,
          "lo": 0.23219210586793973,
          "hi": 0.23809377671631937
        },
        "auroc": {
          "value": 0.6765497423561583,
          "lo": 0.6622366769117792,
          "hi": 0.6908041518723789
        },
        "acc@80": {
          "value": 0.6730739658619099,
          "lo": 0.6583250142349811,
          "hi": 0.6852436048190568
        }
      }
    },
    {
      "env": "chaosnli",
      "slice": "all",
      "system": "LFM2.5-VL-450M",
      "metrics": {
        "n": {
          "value": 2763.0,
          "lo": 2732.475,
          "hi": 2798.525
        },
        "ce": {
          "value": 1.338092801429914,
          "lo": 1.3015494891945418,
          "hi": 1.3777673557125674
        },
        "kl": {
          "value": 0.8077327313606523,
          "lo": 0.7694633892661862,
          "hi": 0.8466904649156057
        },
        "jsd": {
          "value": 0.15568317018900518,
          "lo": 0.14978620375965646,
          "hi": 0.16185003866414396
        },
        "brier": {
          "value": 0.45064982413103327,
          "lo": 0.4318968382561482,
          "hi": 0.4702720495376779
        },
        "acc": {
          "value": 0.47050307636626854,
          "lo": 0.45173159322760664,
          "hi": 0.49106711293539973
        },
        "conf": {
          "value": 0.793341954419719,
          "lo": 0.7883808439698059,
          "hi": 0.7978205623973785
        },
        "ece": {
          "value": 0.32283887805345046,
          "lo": 0.30244659738864715,
          "hi": 0.3422140118566659
        },
        "reliability": {
          "value": 0.11115614403633266,
          "lo": 0.09952787023352347,
          "hi": 0.12475005518126144
        },
        "resolution": {
          "value": 0.004140842407372943,
          "lo": 0.0030534059588853087,
          "hi": 0.007733036003736438
        },
        "uncertainty": {
          "value": 0.24912993149614582,
          "lo": 0.2476698559729648,
          "hi": 0.24992017489438081
        },
        "auroc": {
          "value": 0.5560857563489142,
          "lo": 0.5350635361967264,
          "hi": 0.5771278261325536
        },
        "acc@80": {
          "value": 0.4850678733031674,
          "lo": 0.4625494747149383,
          "hi": 0.5074677074352085
        }
      }
    },
    {
      "env": "chaosnli",
      "slice": "all",
      "system": "LFM2.5-VL-450M +T",
      "metrics": {
        "n": {
          "value": 2763.0,
          "lo": 2732.475,
          "hi": 2798.525
        },
        "ce": {
          "value": 0.9470154748727558,
          "lo": 0.9377579590201713,
          "hi": 0.9565186291632289
        },
        "kl": {
          "value": 0.41665540480349406,
          "lo": 0.4075258234291402,
          "hi": 0.4265268103403057
        },
        "jsd": {
          "value": 0.11686510490943085,
          "lo": 0.11437385167077567,
          "hi": 0.11960367577491623
        },
        "brier": {
          "value": 0.2754697165443986,
          "lo": 0.2682097526227291,
          "hi": 0.283347714769842
        },
        "acc": {
          "value": 0.47050307636626854,
          "lo": 0.45173159322760664,
          "hi": 0.49106711293539973
        },
        "conf": {
          "value": 0.4756218893714501,
          "lo": 0.4715470519749072,
          "hi": 0.4795668726741519
        },
        "ece": {
          "value": 0.060558357153253424,
          "lo": 0.05080237045491872,
          "hi": 0.08167849404299185
        },
        "reliability": {
          "value": 0.005019142385552081,
          "lo": 0.0037667096489379685,
          "hi": 0.009072622700655618
        },
        "resolution": {
          "value": 0.004701090799118026,
          "lo": 0.0034839186490559088,
          "hi": 0.008955117248874217
        },
        "uncertainty": {
          "value": 0.24912993149614582,
          "lo": 0.2476698559729648,
          "hi": 0.24992017489438081
        },
        "auroc": {
          "value": 0.5551511646248488,
          "lo": 0.5348637119219112,
          "hi": 0.5765463281844294
        },
        "acc@80": {
          "value": 0.48733031674208144,
          "lo": 0.46429486455981944,
          "hi": 0.5089358074753532
        }
      }
    },
    {
      "env": "civilcomments",
      "slice": "all",
      "system": "LFM2.5-VL-450M",
      "metrics": {
        "n": {
          "value": 10581.0,
          "lo": 10261.425,
          "hi": 10904.525
        },
        "ce": {
          "value": 1.4821756429077173,
          "lo": 1.4715134964963283,
          "hi": 1.4919811813196462
        },
        "kl": {
          "value": 1.250692834138376,
          "lo": 1.234751864466174,
          "hi": 1.2657587399506547
        },
        "jsd": {
          "value": 0.3094179506553737,
          "lo": 0.30548679187966254,
          "hi": 0.3131230305948588
        },
        "brier": {
          "value": 0.9676093443593095,
          "lo": 0.9558303989434425,
          "hi": 0.9785138985363868
        },
        "acc": {
          "value": 0.08902750212645308,
          "lo": 0.08389836855315358,
          "hi": 0.09475320608057158
        },
        "conf": {
          "value": 0.8087093981021451,
          "lo": 0.8069045121845806,
          "hi": 0.8105242254083899
        },
        "ece": {
          "value": 0.7196818959756921,
          "lo": 0.7141072087205469,
          "hi": 0.7246391429567834
        },
        "reliability": {
          "value": 0.5196409469902246,
          "lo": 0.5114804735713947,
          "hi": 0.5269996362257217
        },
        "resolution": {
          "value": 0.0014444372335602875,
          "lo": 0.0011473712943158997,
          "hi": 0.0020672249708752147
        },
        "uncertainty": {
          "value": 0.08110160599157748,
          "lo": 0.0768594322623401,
          "hi": 0.08577496996869058
        },
        "auroc": {
          "value": 0.6267196427993231,
          "lo": 0.6090423391336587,
          "hi": 0.6445086359013343
        },
        "acc@80": {
          "value": 0.09982279976373302,
          "lo": 0.09379804174920786,
          "hi": 0.10631157687181092
        }
      }
    },
    {
      "env": "civilcomments",
      "slice": "all",
      "system": "LFM2.5-VL-450M +T",
      "metrics": {
        "n": {
          "value": 10581.0,
          "lo": 10261.425,
          "hi": 10904.525
        },
        "ce": {
          "value": 0.6966061569918292,
          "lo": 0.696547118275331,
          "hi": 0.6966604611465038
        },
        "kl": {
          "value": 0.4651233482224881,
          "lo": 0.45887698441033087,
          "hi": 0.4719685436354903
        },
        "jsd": {
          "value": 0.14202715322565218,
          "lo": 0.1400542553069198,
          "hi": 0.14422450005879134
        },
        "brier": {
          "value": 0.3479509513394226,
          "lo": 0.34365343805139087,
          "hi": 0.35265884466994574
        },
        "acc": {
          "value": 0.08902750212645308,
          "lo": 0.08389836855315358,
          "hi": 0.09475320608057158
        },
        "conf": {
          "value": 0.5025412600611476,
          "lo": 0.502522134743794,
          "hi": 0.502559541257449
        },
        "ece": {
          "value": 0.41351375793469464,
          "lo": 0.4077915773159254,
          "hi": 0.41864092887115295
        },
        "reliability": {
          "value": 0.17238127768567704,
          "lo": 0.16785753907234766,
          "hi": 0.17680060406378054
        },
        "resolution": {
          "value": 0.0014444372335602875,
          "lo": 0.0011473712943158997,
          "hi": 0.0020672249708752147
        },
        "uncertainty": {
          "value": 0.08110160599157748,
          "lo": 0.0768594322623401,
          "hi": 0.08577496996869058
        },
        "auroc": {
          "value": 0.6267196427993231,
          "lo": 0.6090423391336587,
          "hi": 0.6445086359013343
        },
        "acc@80": {
          "value": 0.09982279976373302,
          "lo": 0.09379804174920786,
          "hi": 0.10631157687181092
        }
      }
    },
    {
      "env": "folktables",
      "slice": "all",
      "system": "LFM2.5-VL-450M",
      "metrics": {
        "n": {
          "value": 15003.0,
          "lo": 14855.8,
          "hi": 15157.0
        },
        "ce": {
          "value": 1.13187423963659,
          "lo": 1.1199427317069248,
          "hi": 1.143924365145015
        },
        "kl": {
          "value": 1.13187423963659,
          "lo": 1.1199427317069248,
          "hi": 1.143924365145015
        },
        "jsd": {
          "value": 0.2976529663502538,
          "lo": 0.2947501180040379,
          "hi": 0.30064246013239343
        },
        "brier": {
          "value": 0.8518198364734217,
          "lo": 0.8421069098053824,
          "hi": 0.8617736268712928
        },
        "acc": {
          "value": 0.35112977404519097,
          "lo": 0.3427286002698279,
          "hi": 0.3591625122920405
        },
        "conf": {
          "value": 0.8052857982343986,
          "lo": 0.8038829735505919,
          "hi": 0.8066028173974609
        },
        "ece": {
          "value": 0.4541560241892077,
          "lo": 0.44676883916665083,
          "hi": 0.46226413029702884
        },
        "reliability": {
          "value": 0.2133990714594856,
          "lo": 0.2065251674018482,
          "hi": 0.2212758548271539
        },
        "resolution": {
          "value": 0.01625963369096802,
          "lo": 0.014593422216554746,
          "hi": 0.01834318497783375
        },
        "uncertainty": {
          "value": 0.2278376558241641,
          "lo": 0.2252656850700038,
          "hi": 0.23016480033951636
        },
        "auroc": {
          "value": 0.6484255122164856,
          "lo": 0.6393501068451605,
          "hi": 0.6577473978427627
        },
        "acc@80": {
          "value": 0.3798533577737044,
          "lo": 0.37053804935340834,
          "hi": 0.3899598578179518
        }
      }
    },
    {
      "env": "folktables",
      "slice": "all",
      "system": "LFM2.5-VL-450M +T",
      "metrics": {
        "n": {
          "value": 15003.0,
          "lo": 14855.8,
          "hi": 15157.0
        },
        "ce": {
          "value": 0.69433492276209,
          "lo": 0.6942494606398926,
          "hi": 0.6944197895761097
        },
        "kl": {
          "value": 0.69433492276209,
          "lo": 0.6942494606398926,
          "hi": 0.6944197895761097
        },
        "jsd": {
          "value": 0.21608632861145755,
          "lo": 0.21606285112182436,
          "hi": 0.21610963849604864
        },
        "brier": {
          "value": 0.5011877355086589,
          "lo": 0.5011022745545566,
          "hi": 0.5012726015528075
        },
        "acc": {
          "value": 0.35112977404519097,
          "lo": 0.3427286002698279,
          "hi": 0.3591625122920405
        },
        "conf": {
          "value": 0.5024821010314376,
          "lo": 0.5024680409691173,
          "hi": 0.5024953831742714
        },
        "ece": {
          "value": 0.17631529944337335,
          "lo": 0.1679947607190661,
          "hi": 0.18368582074778786
        },
        "reliability": {
          "value": 0.0390112009131181,
          "lo": 0.03597731804506837,
          "hi": 0.04218598137939827
        },
        "resolution": {
          "value": 0.01625963369096802,
          "lo": 0.014593422216554746,
          "hi": 0.01834318497783375
        },
        "uncertainty": {
          "value": 0.2278376558241641,
          "lo": 0.2252656850700038,
          "hi": 0.23016480033951636
        },
        "auroc": {
          "value": 0.6484255122164856,
          "lo": 0.6393501068451605,
          "hi": 0.6577473978427627
        },
        "acc@80": {
          "value": 0.3798533577737044,
          "lo": 0.37053804935340834,
          "hi": 0.3899598578179518
        }
      }
    },
    {
      "env": "mcq",
      "slice": "all",
      "system": "LFM2.5-VL-450M",
      "metrics": {
        "n": {
          "value": 5064.0,
          "lo": 5064.0,
          "hi": 5064.0
        },
        "ce": {
          "value": 1.1238747937701299,
          "lo": 1.0841670112525117,
          "hi": 1.1706346398444945
        },
        "kl": {
          "value": 1.1238747937701299,
          "lo": 1.0841670112525117,
          "hi": 1.1706346398444945
        },
        "jsd": {
          "value": 0.22378367348235934,
          "lo": 0.21775409032394605,
          "hi": 0.23094043551278276
        },
        "brier": {
          "value": 0.53661718405505,
          "lo": 0.5205883125570585,
          "hi": 0.5559808042874332
        },
        "acc": {
          "value": 0.6330963665086888,
          "lo": 0.6201471169036334,
          "hi": 0.6453396524486572
        },
        "conf": {
          "value": 0.8085802885546315,
          "lo": 0.8031755723428469,
          "hi": 0.8137682487780602
        },
        "ece": {
          "value": 0.17548392204594274,
          "lo": 0.16367946177070716,
          "hi": 0.18919804894349182
        },
        "reliability": {
          "value": 0.03901419698912007,
          "lo": 0.03459275241008605,
          "hi": 0.045940238397651
        },
        "resolution": {
          "value": 0.04566062336067558,
          "lo": 0.04170762143482616,
          "hi": 0.051254038565977274
        },
        "uncertainty": {
          "value": 0.23228535722218477,
          "lo": 0.22887638542610353,
          "hi": 0.23556458277959963
        },
        "auroc": {
          "value": 0.7638296936516368,
          "lo": 0.7515525862224941,
          "hi": 0.7775176429487639
        },
        "acc@80": {
          "value": 0.7018020241915577,
          "lo": 0.6856023204147124,
          "hi": 0.7147679585287584
        }
      }
    },
    {
      "env": "mcq",
      "slice": "all",
      "system": "LFM2.5-VL-450M +T",
      "metrics": {
        "n": {
          "value": 5064.0,
          "lo": 5064.0,
          "hi": 5064.0
        },
        "ce": {
          "value": 0.9293923372332724,
          "lo": 0.9085442498482007,
          "hi": 0.9541396485598952
        },
        "kl": {
          "value": 0.9293923372332724,
          "lo": 0.9085442498482007,
          "hi": 0.9541396485598952
        },
        "jsd": {
          "value": 0.24308830993178895,
          "lo": 0.23876913010109363,
          "hi": 0.24814049221299603
        },
        "brier": {
          "value": 0.483246252779501,
          "lo": 0.4721787294287204,
          "hi": 0.49679611510778227
        },
        "acc": {
          "value": 0.6330963665086888,
          "lo": 0.6201471169036334,
          "hi": 0.6453396524486572
        },
        "conf": {
          "value": 0.6332911854783634,
          "lo": 0.6279506724075,
          "hi": 0.6386419969695399
        },
        "ece": {
          "value": 0.025710280334627914,
          "lo": 0.022769106257303048,
          "hi": 0.041568619142218766
        },
        "reliability": {
          "value": 0.0009101891355786433,
          "lo": 0.0008029880638276326,
          "hi": 0.0025178788028172865
        },
        "resolution": {
          "value": 0.047357823574182736,
          "lo": 0.04354459092559669,
          "hi": 0.05275982801643855
        },
        "uncertainty": {
          "value": 0.23228535722218477,
          "lo": 0.22887638542610353,
          "hi": 0.23556458277959963
        },
        "auroc": {
          "value": 0.7685087567914574,
          "lo": 0.7560508452734865,
          "hi": 0.7819789848383302
        },
        "acc@80": {
          "value": 0.7069859294001481,
          "lo": 0.6906813132559863,
          "hi": 0.7193285608491731
        }
      }
    },
    {
      "env": "stsb",
      "slice": "all",
      "system": "LFM2.5-VL-450M",
      "metrics": {
        "n": {
          "value": 3676.0,
          "lo": 3150.95,
          "hi": 4793.674999999996
        },
        "ce": {
          "value": 2.769908074626373,
          "lo": 2.7249642687168563,
          "hi": 2.812541851185826
        },
        "kl": {
          "value": 2.367604251764118,
          "lo": 2.3196694467977057,
          "hi": 2.412967225632331
        },
        "jsd": {
          "value": 0.42961846772895235,
          "lo": 0.41779523612791825,
          "hi": 0.44160509059057923
        },
        "brier": {
          "value": 0.9222239189081299,
          "lo": 0.8803932211717393,
          "hi": 0.9668684382412573
        },
        "acc": {
          "value": 0.23449401523394994,
          "lo": 0.20829456008503736,
          "hi": 0.26186580018791106
        },
        "conf": {
          "value": 0.7663902332428619,
          "lo": 0.7597274453661149,
          "hi": 0.7754193678634801
        },
        "ece": {
          "value": 0.5318962180089121,
          "lo": 0.4991755148723419,
          "hi": 0.5669433273802466
        },
        "reliability": {
          "value": 0.2979379142039276,
          "lo": 0.26392840232646725,
          "hi": 0.3386203519337424
        },
        "resolution": {
          "value": 0.005745888820674186,
          "lo": 0.0037674743621111855,
          "hi": 0.008934000963257147
        },
        "uncertainty": {
          "value": 0.17950657205341,
          "lo": 0.16490790178615253,
          "hi": 0.19329210120204962
        },
        "auroc": {
          "value": 0.3864619972725039,
          "lo": 0.35313031786271104,
          "hi": 0.420721035327209
        },
        "acc@80": {
          "value": 0.20843250595035703,
          "lo": 0.1813403726698036,
          "hi": 0.23840618798127763
        }
      }
    },
    {
      "env": "stsb",
      "slice": "all",
      "system": "LFM2.5-VL-450M +T",
      "metrics": {
        "n": {
          "value": 3676.0,
          "lo": 3150.95,
          "hi": 4793.674999999996
        },
        "ce": {
          "value": 1.7879169759051958,
          "lo": 1.7854613143177502,
          "hi": 1.790315991369925
        },
        "kl": {
          "value": 1.3856131530429403,
          "lo": 1.3744897439790065,
          "hi": 1.3960821204982374
        },
        "jsd": {
          "value": 0.36946552416133693,
          "lo": 0.3673007075324218,
          "hi": 0.3715704740325432
        },
        "brier": {
          "value": 0.557530205458868,
          "lo": 0.5499179112941774,
          "hi": 0.5648913387038841
        },
        "acc": {
          "value": 0.23449401523394994,
          "lo": 0.20829456008503736,
          "hi": 0.26186580018791106
        },
        "conf": {
          "value": 0.1945217882583341,
          "lo": 0.19428222565926243,
          "hi": 0.19481388059383264
        },
        "ece": {
          "value": 0.06466293144934368,
          "lo": 0.05635712184785718,
          "hi": 0.08360482969247139
        },
        "reliability": {
          "value": 0.007829707597464845,
          "lo": 0.005758262436951522,
          "hi": 0.01262360517332117
        },
        "resolution": {
          "value": 0.005915938610681793,
          "lo": 0.00409133744752459,
          "hi": 0.00926977298543524
        },
        "uncertainty": {
          "value": 0.17950657205341,
          "lo": 0.16490790178615253,
          "hi": 0.19329210120204962
        },
        "auroc": {
          "value": 0.3916183088534787,
          "lo": 0.36758265019595104,
          "hi": 0.42022600196576926
        },
        "acc@80": {
          "value": 0.20775246514790888,
          "lo": 0.1796780074561421,
          "hi": 0.23912592842405617
        }
      }
    },
    {
      "env": "yelp",
      "slice": "all",
      "system": "LFM2.5-VL-450M",
      "metrics": {
        "n": {
          "value": 5000.0,
          "lo": 5000.0,
          "hi": 5000.0
        },
        "ce": {
          "value": 2.0268496797689064,
          "lo": 1.9757696337111814,
          "hi": 2.0727043198643322
        },
        "kl": {
          "value": 2.0268496797689064,
          "lo": 1.9757696337111814,
          "hi": 2.0727043198643322
        },
        "jsd": {
          "value": 0.38751170604058394,
          "lo": 0.3800926360793931,
          "hi": 0.3943478975315039
        },
        "brier": {
          "value": 0.9752197802933239,
          "lo": 0.9535710225405619,
          "hi": 0.9955246274883223
        },
        "acc": {
          "value": 0.3654,
          "lo": 0.351295,
          "hi": 0.37961
        },
        "conf": {
          "value": 0.772893846130371,
          "lo": 0.7679695500971377,
          "hi": 0.7770327220417559
        },
        "ece": {
          "value": 0.4074938461303711,
          "lo": 0.3921759519062936,
          "hi": 0.42181115289151666
        },
        "reliability": {
          "value": 0.18062167840384138,
          "lo": 0.16894849842848417,
          "hi": 0.1937890895997626
        },
        "resolution": {
          "value": 0.002703533747639855,
          "lo": 0.0019412275991941928,
          "hi": 0.005002970304404658
        },
        "uncertainty": {
          "value": 0.23188284000000003,
          "lo": 0.22788681300000002,
          "hi": 0.23550620800000002
        },
        "auroc": {
          "value": 0.5561342098449372,
          "lo": 0.540110963525977,
          "hi": 0.5729905033824033
        },
        "acc@80": {
          "value": 0.38675,
          "lo": 0.37110625,
          "hi": 0.40163125
        }
      }
    },
    {
      "env": "yelp",
      "slice": "all",
      "system": "LFM2.5-VL-450M +T",
      "metrics": {
        "n": {
          "value": 5000.0,
          "lo": 5000.0,
          "hi": 5000.0
        },
        "ce": {
          "value": 1.4661425923220464,
          "lo": 1.4517260160875394,
          "hi": 1.4799758798314981
        },
        "kl": {
          "value": 1.4661425923220464,
          "lo": 1.4517260160875394,
          "hi": 1.4799758798314981
        },
        "jsd": {
          "value": 0.38481973653617163,
          "lo": 0.38190879602015265,
          "hi": 0.38753674066981
        },
        "brier": {
          "value": 0.7442832037214115,
          "lo": 0.7368222931464771,
          "hi": 0.7510716412102759
        },
        "acc": {
          "value": 0.3654,
          "lo": 0.351295,
          "hi": 0.37961
        },
        "conf": {
          "value": 0.3929922703588896,
          "lo": 0.3906625462630033,
          "hi": 0.3950085689472873
        },
        "ece": {
          "value": 0.03837137950549759,
          "lo": 0.03302673934354592,
          "hi": 0.05504690977148236
        },
        "reliability": {
          "value": 0.002792990203293114,
          "lo": 0.00206872995279149,
          "hi": 0.004910302827656831
        },
        "resolution": {
          "value": 0.0026951055589721448,
          "lo": 0.0019255613761626623,
          "hi": 0.004816427312298835
        },
        "uncertainty": {
          "value": 0.23188284000000003,
          "lo": 0.22788681300000002,
          "hi": 0.23550620800000002
        },
        "auroc": {
          "value": 0.5551914061428608,
          "lo": 0.5389410313371229,
          "hi": 0.5711703153645804
        },
        "acc@80": {
          "value": 0.387,
          "lo": 0.3715,
          "hi": 0.4027625
        }
      }
    },
    {
      "env": "boolq",
      "slice": "all",
      "system": "zs-LFM2.5-VL-450M-prefill",
      "metrics": {
        "n": {
          "value": 8129.0,
          "lo": 7984.475,
          "hi": 8270.575
        },
        "ce": {
          "value": 2.19728081495343,
          "lo": 2.124304608559749,
          "hi": 2.2815227998032035
        },
        "kl": {
          "value": 2.19728081495343,
          "lo": 2.124304608559749,
          "hi": 2.2815227998032035
        },
        "jsd": {
          "value": 0.25085200285439146,
          "lo": 0.24325014396883182,
          "hi": 0.2596134570244947
        },
        "brier": {
          "value": 0.7293619975390971,
          "lo": 0.7070458137496666,
          "hi": 0.7550569114874577
        },
        "acc": {
          "value": 0.6255381965801452,
          "lo": 0.6123595475426337,
          "hi": 0.6370788578916808
        },
        "conf": {
          "value": 0.9897076468330451,
          "lo": 0.9886793475548706,
          "hi": 0.9906145027240567
        },
        "ece": {
          "value": 0.36416945025289993,
          "lo": 0.3526348168184627,
          "hi": 0.3770581720829012
        },
        "reliability": {
          "value": 0.1483079643460137,
          "lo": 0.1392692926735881,
          "hi": 0.15822142319007584
        },
        "resolution": {
          "value": 0.019428780862266886,
          "lo": 0.016427198057464343,
          "hi": 0.023126909746466903
        },
        "uncertainty": {
          "value": 0.23424016119940483,
          "lo": 0.2312093865609688,
          "hi": 0.237375277593653
        },
        "auroc": {
          "value": 0.6688632601878447,
          "lo": 0.6535178089270295,
          "hi": 0.683143197873873
        },
        "acc@80": {
          "value": 0.6763032446563124,
          "lo": 0.6616667289131235,
          "hi": 0.6883528348868468
        }
      }
    },
    {
      "env": "boolq",
      "slice": "all",
      "system": "zs-LFM2.5-VL-450M-prefill +T",
      "metrics": {
        "n": {
          "value": 8129.0,
          "lo": 7984.475,
          "hi": 8270.575
        },
        "ce": {
          "value": 0.6456473889972213,
          "lo": 0.6386110520713849,
          "hi": 0.6540006891041504
        },
        "kl": {
          "value": 0.6456473889972213,
          "lo": 0.6386110520713849,
          "hi": 0.6540006891041504
        },
        "jsd": {
          "value": 0.1977480080233696,
          "lo": 0.19585256710888305,
          "hi": 0.2000011000174901
        },
        "brier": {
          "value": 0.45409670547516723,
          "lo": 0.44733937324373924,
          "hi": 0.4621110447796479
        },
        "acc": {
          "value": 0.6255381965801452,
          "lo": 0.6123595475426337,
          "hi": 0.6370788578916808
        },
        "conf": {
          "value": 0.6503467435365148,
          "lo": 0.6494268606073704,
          "hi": 0.6511364278149309
        },
        "ece": {
          "value": 0.09484015765890995,
          "lo": 0.08239948310358765,
          "hi": 0.10614650821261719
        },
        "reliability": {
          "value": 0.012131099068849526,
          "lo": 0.009954431487962349,
          "hi": 0.015130793748406441
        },
        "resolution": {
          "value": 0.019428780862266886,
          "lo": 0.016427198057464343,
          "hi": 0.023126909746466903
        },
        "uncertainty": {
          "value": 0.23424016119940483,
          "lo": 0.2312093865609688,
          "hi": 0.237375277593653
        },
        "auroc": {
          "value": 0.6688632601878447,
          "lo": 0.6535178089270295,
          "hi": 0.683143197873873
        },
        "acc@80": {
          "value": 0.6763032446563124,
          "lo": 0.6616667289131235,
          "hi": 0.6883528348868468
        }
      }
    },
    {
      "env": "chaosnli",
      "slice": "all",
      "system": "zs-LFM2.5-VL-450M-prefill",
      "metrics": {
        "n": {
          "value": 2763.0,
          "lo": 2732.475,
          "hi": 2798.525
        },
        "ce": {
          "value": 1.5707520474601657,
          "lo": 1.5275747663629993,
          "hi": 1.6173270858002489
        },
        "kl": {
          "value": 1.0403919773909036,
          "lo": 0.9966209037282945,
          "hi": 1.0856548039288445
        },
        "jsd": {
          "value": 0.17964967086268427,
          "lo": 0.17321191794730745,
          "hi": 0.18619182691973712
        },
        "brier": {
          "value": 0.5303528884034465,
          "lo": 0.5096208612341412,
          "hi": 0.5517010621282801
        },
        "acc": {
          "value": 0.47701773434672456,
          "lo": 0.45927291945657245,
          "hi": 0.4972083924734921
        },
        "conf": {
          "value": 0.8680714446049517,
          "lo": 0.8640764415692318,
          "hi": 0.8719903504246821
        },
        "ece": {
          "value": 0.3910537102582271,
          "lo": 0.36983254462392084,
          "hi": 0.409753591755807
        },
        "reliability": {
          "value": 0.16377605405545875,
          "lo": 0.1487661016563696,
          "hi": 0.1803656095749142
        },
        "resolution": {
          "value": 0.0037965216513373306,
          "lo": 0.0027810791537164403,
          "hi": 0.007693189873688164
        },
        "uncertainty": {
          "value": 0.24947181546544228,
          "lo": 0.24834130451031894,
          "hi": 0.24998980524401404
        },
        "auroc": {
          "value": 0.542554515334653,
          "lo": 0.5215917921541139,
          "hi": 0.5615865777972047
        },
        "acc@80": {
          "value": 0.4832579185520362,
          "lo": 0.4612578412985837,
          "hi": 0.5049284446588672
        }
      }
    },
    {
      "env": "chaosnli",
      "slice": "all",
      "system": "zs-LFM2.5-VL-450M-prefill +T",
      "metrics": {
        "n": {
          "value": 2763.0,
          "lo": 2732.475,
          "hi": 2798.525
        },
        "ce": {
          "value": 0.9411396746035926,
          "lo": 0.9316432723286321,
          "hi": 0.9510239685529941
        },
        "kl": {
          "value": 0.4107796045343307,
          "lo": 0.40118868149631065,
          "hi": 0.42065309138964746
        },
        "jsd": {
          "value": 0.11444784812915831,
          "lo": 0.11182792046593194,
          "hi": 0.11709707186892492
        },
        "brier": {
          "value": 0.2715981745898122,
          "lo": 0.2638461718757901,
          "hi": 0.2790434599477199
        },
        "acc": {
          "value": 0.47701773434672456,
          "lo": 0.45927291945657245,
          "hi": 0.4972083924734921
        },
        "conf": {
          "value": 0.4964434871110875,
          "lo": 0.4928107712749071,
          "hi": 0.5000973279827943
        },
        "ece": {
          "value": 0.04567342424405743,
          "lo": 0.039554822050559726,
          "hi": 0.0703830773513986
        },
        "reliability": {
          "value": 0.003249705978136044,
          "lo": 0.0024085959787744924,
          "hi": 0.007313038874674322
        },
        "resolution": {
          "value": 0.006150872707803075,
          "lo": 0.004592221659527366,
          "hi": 0.010323172091024698
        },
        "uncertainty": {
          "value": 0.24947181546544228,
          "lo": 0.24834130451031894,
          "hi": 0.24998980524401404
        },
        "auroc": {
          "value": 0.5787131598153856,
          "lo": 0.558184534930305,
          "hi": 0.5992775062158711
        },
        "acc@80": {
          "value": 0.5027149321266968,
          "lo": 0.48357814607970573,
          "hi": 0.5246717077591531
        }
      }
    },
    {
      "env": "civilcomments",
      "slice": "all",
      "system": "zs-LFM2.5-VL-450M-prefill",
      "metrics": {
        "n": {
          "value": 10581.0,
          "lo": 10261.425,
          "hi": 10904.525
        },
        "ce": {
          "value": 0.5900203309821959,
          "lo": 0.5838787321375208,
          "hi": 0.5954055604731321
        },
        "kl": {
          "value": 0.3585375222128549,
          "lo": 0.3526599668328442,
          "hi": 0.3646889333871
        },
        "jsd": {
          "value": 0.10899092024426695,
          "lo": 0.10722699098880814,
          "hi": 0.11086045531598936
        },
        "brier": {
          "value": 0.24589472982697438,
          "lo": 0.24149032242307591,
          "hi": 0.25058992920458706
        },
        "acc": {
          "value": 0.7690199414044041,
          "lo": 0.7604835999526309,
          "hi": 0.7782208280198362
        },
        "conf": {
          "value": 0.6458412299376108,
          "lo": 0.643821715178355,
          "hi": 0.6479629327616658
        },
        "ece": {
          "value": 0.1231787114667933,
          "lo": 0.11508685145064275,
          "hi": 0.13185081974310223
        },
        "reliability": {
          "value": 0.01598794750263893,
          "lo": 0.01415087215041063,
          "hi": 0.018437871557114755
        },
        "resolution": {
          "value": 0.005509658827805386,
          "lo": 0.00455771339467265,
          "hi": 0.006841205371086549
        },
        "uncertainty": {
          "value": 0.177628271126771,
          "lo": 0.17259316108493603,
          "hi": 0.18214829408492678
        },
        "auroc": {
          "value": 0.6114104019001925,
          "lo": 0.5994015242532421,
          "hi": 0.6240694745868953
        },
        "acc@80": {
          "value": 0.7942114589486119,
          "lo": 0.7824591945486796,
          "hi": 0.8012385112850875
        }
      }
    },
    {
      "env": "civilcomments",
      "slice": "all",
      "system": "zs-LFM2.5-VL-450M-prefill +T",
      "metrics": {
        "n": {
          "value": 10581.0,
          "lo": 10261.425,
          "hi": 10904.525
        },
        "ce": {
          "value": 0.5814737265683071,
          "lo": 0.5732353596092621,
          "hi": 0.5890529160665945
        },
        "kl": {
          "value": 0.3499909177989659,
          "lo": 0.3433920464430891,
          "hi": 0.35686043475600315
        },
        "jsd": {
          "value": 0.10305108396492324,
          "lo": 0.10123033439323824,
          "hi": 0.10503232184467196
        },
        "brier": {
          "value": 0.2374545346605293,
          "lo": 0.2319786130597003,
          "hi": 0.24301453536356327
        },
        "acc": {
          "value": 0.7690199414044041,
          "lo": 0.7604835999526309,
          "hi": 0.7782208280198362
        },
        "conf": {
          "value": 0.6966262697833534,
          "lo": 0.6940898809225147,
          "hi": 0.6992943953368023
        },
        "ece": {
          "value": 0.07537646986127365,
          "lo": 0.06830649884157658,
          "hi": 0.08475949545171395
        },
        "reliability": {
          "value": 0.007450847038259102,
          "lo": 0.006266291426017038,
          "hi": 0.009229124500393811
        },
        "resolution": {
          "value": 0.005509658827805386,
          "lo": 0.00455771339467265,
          "hi": 0.006841205371086549
        },
        "uncertainty": {
          "value": 0.177628271126771,
          "lo": 0.17259316108493603,
          "hi": 0.18214829408492678
        },
        "auroc": {
          "value": 0.6114104019001925,
          "lo": 0.5994015242532421,
          "hi": 0.6240694745868953
        },
        "acc@80": {
          "value": 0.7942114589486119,
          "lo": 0.7824591945486796,
          "hi": 0.8012385112850875
        }
      }
    },
    {
      "env": "folktables",
      "slice": "all",
      "system": "zs-LFM2.5-VL-450M-prefill",
      "metrics": {
        "n": {
          "value": 15003.0,
          "lo": 14855.8,
          "hi": 15157.0
        },
        "ce": {
          "value": 0.8381660962858019,
          "lo": 0.8311775782412278,
          "hi": 0.8453279801409322
        },
        "kl": {
          "value": 0.8381660962858019,
          "lo": 0.8311775782412278,
          "hi": 0.8453279801409322
        },
        "jsd": {
          "value": 0.24380680319635128,
          "lo": 0.24198023830936183,
          "hi": 0.24565457418496592
        },
        "brier": {
          "value": 0.6295013098606138,
          "lo": 0.6233081387799438,
          "hi": 0.6359303666896823
        },
        "acc": {
          "value": 0.42118243018063056,
          "lo": 0.41307011683913974,
          "hi": 0.4287435405329441
        },
        "conf": {
          "value": 0.6760865042251419,
          "lo": 0.6744624158494803,
          "hi": 0.6777140352172191
        },
        "ece": {
          "value": 0.25705756810492836,
          "lo": 0.24951568640967292,
          "hi": 0.26552571260322566
        },
        "reliability": {
          "value": 0.07892091224121713,
          "lo": 0.07498311484665345,
          "hi": 0.08387523515382338
        },
        "resolution": {
          "value": 0.008360669342411055,
          "lo": 0.007178693961924451,
          "hi": 0.010094067568160319
        },
        "uncertainty": {
          "value": 0.2437877906877688,
          "lo": 0.2424431754297249,
          "hi": 0.24492251689546507
        },
        "auroc": {
          "value": 0.5193538780959998,
          "lo": 0.5098087090592502,
          "hi": 0.5285425283334804
        },
        "acc@80": {
          "value": 0.4035160806532245,
          "lo": 0.3946298976790928,
          "hi": 0.41245959992217585
        }
      }
    },
    {
      "env": "folktables",
      "slice": "all",
      "system": "zs-LFM2.5-VL-450M-prefill +T",
      "metrics": {
        "n": {
          "value": 15003.0,
          "lo": 14855.8,
          "hi": 15157.0
        },
        "ce": {
          "value": 0.6934844593415295,
          "lo": 0.6934336599007378,
          "hi": 0.693534520171246
        },
        "kl": {
          "value": 0.6934844593415295,
          "lo": 0.6934336599007378,
          "hi": 0.693534520171246
        },
        "jsd": {
          "value": 0.21585368991239473,
          "lo": 0.21583973124532319,
          "hi": 0.21586743937159877
        },
        "brier": {
          "value": 0.5003372788995907,
          "lo": 0.5002864798773798,
          "hi": 0.5003873393791819
        },
        "acc": {
          "value": 0.42118243018063056,
          "lo": 0.41307011683913974,
          "hi": 0.4287435405329441
        },
        "conf": {
          "value": 0.5013043170496617,
          "lo": 0.5012913479575357,
          "hi": 0.5013185173855981
        },
        "ece": {
          "value": 0.11145470539128079,
          "lo": 0.10300510627501262,
          "hi": 0.12032274662126989
        },
        "reliability": {
          "value": 0.014739607237335837,
          "lo": 0.012980927601164585,
          "hi": 0.017113136480238565
        },
        "resolution": {
          "value": 0.008360669342411055,
          "lo": 0.007178693961924451,
          "hi": 0.010094067568160319
        },
        "uncertainty": {
          "value": 0.2437877906877688,
          "lo": 0.2424431754297249,
          "hi": 0.24492251689546507
        },
        "auroc": {
          "value": 0.5193538780959998,
          "lo": 0.5098087090592502,
          "hi": 0.5285425283334804
        },
        "acc@80": {
          "value": 0.4035160806532245,
          "lo": 0.3946298976790928,
          "hi": 0.41245959992217585
        }
      }
    },
    {
      "env": "mcq",
      "slice": "all",
      "system": "zs-LFM2.5-VL-450M-prefill",
      "metrics": {
        "n": {
          "value": 5064.0,
          "lo": 5064.0,
          "hi": 5064.0
        },
        "ce": {
          "value": 1.2006815423871746,
          "lo": 1.1556908086451052,
          "hi": 1.2525439528494509
        },
        "kl": {
          "value": 1.2006815423871746,
          "lo": 1.1556908086451052,
          "hi": 1.2525439528494509
        },
        "jsd": {
          "value": 0.22489259461281508,
          "lo": 0.21893657284624224,
          "hi": 0.23211634280046434
        },
        "brier": {
          "value": 0.5490602447358751,
          "lo": 0.5313253217366278,
          "hi": 0.5696447789893158
        },
        "acc": {
          "value": 0.6317140600315956,
          "lo": 0.618088467614534,
          "hi": 0.6442584913112164
        },
        "conf": {
          "value": 0.8235576231368062,
          "lo": 0.8179018961868909,
          "hi": 0.8288325505291434
        },
        "ece": {
          "value": 0.1918435631052107,
          "lo": 0.1803097437399235,
          "hi": 0.20564171560182226
        },
        "reliability": {
          "value": 0.04628672012359426,
          "lo": 0.041624676116715796,
          "hi": 0.05335272961327013
        },
        "resolution": {
          "value": 0.04594570047395249,
          "lo": 0.04148231527304997,
          "hi": 0.05110003801118209
        },
        "uncertainty": {
          "value": 0.23265140638999324,
          "lo": 0.22918947796015116,
          "hi": 0.23605511381645117
        },
        "auroc": {
          "value": 0.7650734520757576,
          "lo": 0.7519250235530622,
          "hi": 0.7782467152900953
        },
        "acc@80": {
          "value": 0.6998272031597137,
          "lo": 0.6853801530486299,
          "hi": 0.7143915082695631
        }
      }
    },
    {
      "env": "mcq",
      "slice": "all",
      "system": "zs-LFM2.5-VL-450M-prefill +T",
      "metrics": {
        "n": {
          "value": 5064.0,
          "lo": 5064.0,
          "hi": 5064.0
        },
        "ce": {
          "value": 0.9334720356419071,
          "lo": 0.914514870062201,
          "hi": 0.9582322834004864
        },
        "kl": {
          "value": 0.9334720356419071,
          "lo": 0.914514870062201,
          "hi": 0.9582322834004864
        },
        "jsd": {
          "value": 0.24389875313034248,
          "lo": 0.23958877153740263,
          "hi": 0.24891649684701952
        },
        "brier": {
          "value": 0.4846589365397375,
          "lo": 0.47352779484961377,
          "hi": 0.4980496994037511
        },
        "acc": {
          "value": 0.6317140600315956,
          "lo": 0.618088467614534,
          "hi": 0.6442584913112164
        },
        "conf": {
          "value": 0.6321695269208738,
          "lo": 0.6264940350916997,
          "hi": 0.6378483184065368
        },
        "ece": {
          "value": 0.022256455628038435,
          "lo": 0.01999505672676786,
          "hi": 0.039566132825991016
        },
        "reliability": {
          "value": 0.0007071546382037909,
          "lo": 0.000646030699779827,
          "hi": 0.002216121494060057
        },
        "resolution": {
          "value": 0.046863271490700284,
          "lo": 0.04270781407907646,
          "hi": 0.05233557744229133
        },
        "uncertainty": {
          "value": 0.23265140638999324,
          "lo": 0.22918947796015116,
          "hi": 0.23605511381645117
        },
        "auroc": {
          "value": 0.7688532860889001,
          "lo": 0.7557556679608567,
          "hi": 0.782075236828284
        },
        "acc@80": {
          "value": 0.7010614663046162,
          "lo": 0.6855097506788447,
          "hi": 0.7151320661565046
        }
      }
    },
    {
      "env": "stsb",
      "slice": "all",
      "system": "zs-LFM2.5-VL-450M-prefill",
      "metrics": {
        "n": {
          "value": 3676.0,
          "lo": 3150.95,
          "hi": 4793.674999999996
        },
        "ce": {
          "value": 2.9252958890973293,
          "lo": 2.8542557464838847,
          "hi": 2.990760903901766
        },
        "kl": {
          "value": 2.5229920662350733,
          "lo": 2.4474793436751017,
          "hi": 2.5950072765919248
        },
        "jsd": {
          "value": 0.4405102692623737,
          "lo": 0.42611439867612205,
          "hi": 0.45603016319697026
        },
        "brier": {
          "value": 1.0110556365336225,
          "lo": 0.9646760809995163,
          "hi": 1.061952511853829
        },
        "acc": {
          "value": 0.23449401523394994,
          "lo": 0.20829456008503736,
          "hi": 0.26186580018791106
        },
        "conf": {
          "value": 0.8349484990840379,
          "lo": 0.8289625878220371,
          "hi": 0.8429096477418141
        },
        "ece": {
          "value": 0.6004544838500879,
          "lo": 0.5682166316242653,
          "hi": 0.6344356695474724
        },
        "reliability": {
          "value": 0.3760320688752832,
          "lo": 0.33785596690004494,
          "hi": 0.42015100287903556
        },
        "resolution": {
          "value": 0.006868195834272034,
          "lo": 0.004344999612472737,
          "hi": 0.009792027695906771
        },
        "uncertainty": {
          "value": 0.17950657205341,
          "lo": 0.16490790178615253,
          "hi": 0.19329210120204962
        },
        "auroc": {
          "value": 0.3793089161418628,
          "lo": 0.3468237587038948,
          "hi": 0.4127735986935907
        },
        "acc@80": {
          "value": 0.20707242434546072,
          "lo": 0.17569723088852932,
          "hi": 0.23624905508074437
        }
      }
    },
    {
      "env": "stsb",
      "slice": "all",
      "system": "zs-LFM2.5-VL-450M-prefill +T",
      "metrics": {
        "n": {
          "value": 3676.0,
          "lo": 3150.95,
          "hi": 4793.674999999996
        },
        "ce": {
          "value": 1.7818855315091495,
          "lo": 1.776248229095906,
          "hi": 1.786696177576579
        },
        "kl": {
          "value": 1.379581708646894,
          "lo": 1.36617426785534,
          "hi": 1.3919406261312177
        },
        "jsd": {
          "value": 0.36735733088340217,
          "lo": 0.36466329704064493,
          "hi": 0.3699104853924312
        },
        "brier": {
          "value": 0.5556155232652956,
          "lo": 0.5470766212231325,
          "hi": 0.5634638237610552
        },
        "acc": {
          "value": 0.23449401523394994,
          "lo": 0.20829456008503736,
          "hi": 0.26186580018791106
        },
        "conf": {
          "value": 0.21819528485698236,
          "lo": 0.21767100327886024,
          "hi": 0.21884265661817287
        },
        "ece": {
          "value": 0.06393672301189443,
          "lo": 0.05676741520385015,
          "hi": 0.08194139559936113
        },
        "reliability": {
          "value": 0.007255352947775862,
          "lo": 0.005509254823106484,
          "hi": 0.01145279891547225
        },
        "resolution": {
          "value": 0.006273143829640579,
          "lo": 0.004612689575726061,
          "hi": 0.009656098696653346
        },
        "uncertainty": {
          "value": 0.17950657205341,
          "lo": 0.16490790178615253,
          "hi": 0.19329210120204962
        },
        "auroc": {
          "value": 0.38472659902344425,
          "lo": 0.35672093422911605,
          "hi": 0.4155279687971298
        },
        "acc@80": {
          "value": 0.20877252635158108,
          "lo": 0.1788658812922375,
          "hi": 0.2373745525689766
        }
      }
    },
    {
      "env": "yelp",
      "slice": "all",
      "system": "zs-LFM2.5-VL-450M-prefill",
      "metrics": {
        "n": {
          "value": 5000.0,
          "lo": 5000.0,
          "hi": 5000.0
        },
        "ce": {
          "value": 1.909529184309992,
          "lo": 1.8689111656271569,
          "hi": 1.947252798147701
        },
        "kl": {
          "value": 1.909529184309992,
          "lo": 1.8689111656271569,
          "hi": 1.947252798147701
        },
        "jsd": {
          "value": 0.3865241606078184,
          "lo": 0.3793586320580127,
          "hi": 0.3933433205318795
        },
        "brier": {
          "value": 0.9623211190083816,
          "lo": 0.9417940605110796,
          "hi": 0.9803091733259967
        },
        "acc": {
          "value": 0.3508,
          "lo": 0.336295,
          "hi": 0.3652
        },
        "conf": {
          "value": 0.7536129065871239,
          "lo": 0.7485283327189088,
          "hi": 0.7582657927136123
        },
        "ece": {
          "value": 0.4028129065871239,
          "lo": 0.3893723001004755,
          "hi": 0.41627636068210006
        },
        "reliability": {
          "value": 0.17593504065367982,
          "lo": 0.16507473345946477,
          "hi": 0.18765280021593223
        },
        "resolution": {
          "value": 0.004805350649765269,
          "lo": 0.0038690200729840376,
          "hi": 0.007448485908352989
        },
        "uncertainty": {
          "value": 0.22773936,
          "lo": 0.22320066299999997,
          "hi": 0.23182896000000003
        },
        "auroc": {
          "value": 0.5829090764108584,
          "lo": 0.5658942332599112,
          "hi": 0.599821896088401
        },
        "acc@80": {
          "value": 0.37275,
          "lo": 0.35911875,
          "hi": 0.38938125
        }
      }
    },
    {
      "env": "yelp",
      "slice": "all",
      "system": "zs-LFM2.5-VL-450M-prefill +T",
      "metrics": {
        "n": {
          "value": 5000.0,
          "lo": 5000.0,
          "hi": 5000.0
        },
        "ce": {
          "value": 1.4546720351361775,
          "lo": 1.4407163563912457,
          "hi": 1.4679473982561566
        },
        "kl": {
          "value": 1.4546720351361775,
          "lo": 1.4407163563912457,
          "hi": 1.4679473982561566
        },
        "jsd": {
          "value": 0.3824816280037077,
          "lo": 0.3795358432372821,
          "hi": 0.3852028864033026
        },
        "brier": {
          "value": 0.7421864816303968,
          "lo": 0.734649832104407,
          "hi": 0.7490823032617606
        },
        "acc": {
          "value": 0.3508,
          "lo": 0.336295,
          "hi": 0.3652
        },
        "conf": {
          "value": 0.4019106739742869,
          "lo": 0.3991919691836869,
          "hi": 0.40414484885959845
        },
        "ece": {
          "value": 0.05146108965461193,
          "lo": 0.042050024498915935,
          "hi": 0.06641135791941125
        },
        "reliability": {
          "value": 0.0035359386025178735,
          "lo": 0.00256163157206236,
          "hi": 0.006047043118079161
        },
        "resolution": {
          "value": 0.0047180908266350186,
          "lo": 0.0036765392781320312,
          "hi": 0.0073615620756998425
        },
        "uncertainty": {
          "value": 0.22773936,
          "lo": 0.22320066299999997,
          "hi": 0.23182896000000003
        },
        "auroc": {
          "value": 0.5833198091010706,
          "lo": 0.5654694588668958,
          "hi": 0.6004176415078453
        },
        "acc@80": {
          "value": 0.376,
          "lo": 0.36136875,
          "hi": 0.3925
        }
      }
    },
    {
      "env": "boolq",
      "slice": "all",
      "system": "LFM2.5-VL-1.6B",
      "metrics": {
        "n": {
          "value": 8129.0,
          "lo": 7984.475,
          "hi": 8270.575
        },
        "ce": {
          "value": 0.49785330640044806,
          "lo": 0.47560878566083276,
          "hi": 0.5168941670796923
        },
        "kl": {
          "value": 0.49785330640044806,
          "lo": 0.47560878566083276,
          "hi": 0.5168941670796923
        },
        "jsd": {
          "value": 0.1289900321502576,
          "lo": 0.12450703835114382,
          "hi": 0.13344787634126143
        },
        "brier": {
          "value": 0.31364132154159263,
          "lo": 0.2992261822655051,
          "hi": 0.32655412303134806
        },
        "acc": {
          "value": 0.7834912043301759,
          "lo": 0.7727635141427726,
          "hi": 0.7942986561315352
        },
        "conf": {
          "value": 0.8385633099546935,
          "lo": 0.8342370984319535,
          "hi": 0.8421433070067155
        },
        "ece": {
          "value": 0.05572035842341299,
          "lo": 0.04829541495473734,
          "hi": 0.06622252562054536
        },
        "reliability": {
          "value": 0.004336838053223081,
          "lo": 0.003321665032817823,
          "hi": 0.0059174203416841745
        },
        "resolution": {
          "value": 0.017281614936288374,
          "lo": 0.015054374950629739,
          "hi": 0.019808555621349623
        },
        "uncertainty": {
          "value": 0.16963273706742643,
          "lo": 0.16338830068774005,
          "hi": 0.17559999038847476
        },
        "auroc": {
          "value": 0.720295839934912,
          "lo": 0.7059285762321004,
          "hi": 0.73328066474667
        },
        "acc@80": {
          "value": 0.8342303552206674,
          "lo": 0.8219057567913703,
          "hi": 0.8448184058249903
        }
      }
    },
    {
      "env": "boolq",
      "slice": "all",
      "system": "LFM2.5-VL-1.6B +T",
      "metrics": {
        "n": {
          "value": 8129.0,
          "lo": 7984.475,
          "hi": 8270.575
        },
        "ce": {
          "value": 0.47054909267913464,
          "lo": 0.4549301610766264,
          "hi": 0.48505115783006075
        },
        "kl": {
          "value": 0.47054909267913464,
          "lo": 0.4549301610766264,
          "hi": 0.48505115783006075
        },
        "jsd": {
          "value": 0.134867241463773,
          "lo": 0.13130031093067707,
          "hi": 0.1385972596511396
        },
        "brier": {
          "value": 0.3060432354808317,
          "lo": 0.29394893888705187,
          "hi": 0.31721282900461956
        },
        "acc": {
          "value": 0.7834912043301759,
          "lo": 0.7727635141427726,
          "hi": 0.7942986561315352
        },
        "conf": {
          "value": 0.7873545146192261,
          "lo": 0.7830510087868503,
          "hi": 0.7909054094918448
        },
        "ece": {
          "value": 0.02205472772536668,
          "lo": 0.017562225471831002,
          "hi": 0.03129979479071034
        },
        "reliability": {
          "value": 0.0006079931936090638,
          "lo": 0.0004712707639423926,
          "hi": 0.001431221786267658
        },
        "resolution": {
          "value": 0.017281614936288374,
          "lo": 0.015054374950629739,
          "hi": 0.019808555621349623
        },
        "uncertainty": {
          "value": 0.16963273706742643,
          "lo": 0.16338830068774005,
          "hi": 0.17559999038847476
        },
        "auroc": {
          "value": 0.720295839934912,
          "lo": 0.7059285762321004,
          "hi": 0.73328066474667
        },
        "acc@80": {
          "value": 0.8342303552206674,
          "lo": 0.8219057567913703,
          "hi": 0.8448184058249903
        }
      }
    },
    {
      "env": "chaosnli",
      "slice": "all",
      "system": "LFM2.5-VL-1.6B",
      "metrics": {
        "n": {
          "value": 2763.0,
          "lo": 2732.475,
          "hi": 2798.525
        },
        "ce": {
          "value": 0.9602479808993487,
          "lo": 0.9361369244725344,
          "hi": 0.9843172571819331
        },
        "kl": {
          "value": 0.4298879108300868,
          "lo": 0.41134151220678244,
          "hi": 0.4520748675028292
        },
        "jsd": {
          "value": 0.09649697096993072,
          "lo": 0.09281219393386245,
          "hi": 0.10082829836391946
        },
        "brier": {
          "value": 0.26577383186749354,
          "lo": 0.2546888605608322,
          "hi": 0.27974481953511815
        },
        "acc": {
          "value": 0.579080709373869,
          "lo": 0.5594891914460797,
          "hi": 0.5986598971194417
        },
        "conf": {
          "value": 0.7066341453904998,
          "lo": 0.7019352197997519,
          "hi": 0.7123606447036466
        },
        "ece": {
          "value": 0.12755343601663088,
          "lo": 0.11026264951346527,
          "hi": 0.14841342371534058
        },
        "reliability": {
          "value": 0.019211434563092787,
          "lo": 0.015482553590135497,
          "hi": 0.026132798028573642
        },
        "resolution": {
          "value": 0.011672201689484832,
          "lo": 0.009373052472500086,
          "hi": 0.016802626077061344
        },
        "uncertainty": {
          "value": 0.24374624140492565,
          "lo": 0.24026622284850968,
          "hi": 0.24646102634681313
        },
        "auroc": {
          "value": 0.6226829858125538,
          "lo": 0.6014047939769757,
          "hi": 0.6428758549104638
        },
        "acc@80": {
          "value": 0.6131221719457014,
          "lo": 0.5899383468834688,
          "hi": 0.6332811174567478
        }
      }
    },
    {
      "env": "chaosnli",
      "slice": "all",
      "system": "LFM2.5-VL-1.6B +T",
      "metrics": {
        "n": {
          "value": 2763.0,
          "lo": 2732.475,
          "hi": 2798.525
        },
        "ce": {
          "value": 0.8519077402493126,
          "lo": 0.8405353646498452,
          "hi": 0.8627873330566642
        },
        "kl": {
          "value": 0.3215476701800507,
          "lo": 0.3130729840825459,
          "hi": 0.3318221108548847
        },
        "jsd": {
          "value": 0.08853265402241807,
          "lo": 0.08620149330702792,
          "hi": 0.09129035095982928
        },
        "brier": {
          "value": 0.21624319391507632,
          "lo": 0.2094406116514185,
          "hi": 0.22423506366859275
        },
        "acc": {
          "value": 0.579080709373869,
          "lo": 0.5594891914460797,
          "hi": 0.5986598971194417
        },
        "conf": {
          "value": 0.5528968678800066,
          "lo": 0.5493274220417315,
          "hi": 0.5568968168888413
        },
        "ece": {
          "value": 0.04217048857273199,
          "lo": 0.03485165644347747,
          "hi": 0.06340023380791397
        },
        "reliability": {
          "value": 0.0028069758520380406,
          "lo": 0.0019621215165353534,
          "hi": 0.0060221874232393195
        },
        "resolution": {
          "value": 0.011836966536868603,
          "lo": 0.009035809731571719,
          "hi": 0.01629494941176776
        },
        "uncertainty": {
          "value": 0.24374624140492565,
          "lo": 0.24026622284850968,
          "hi": 0.24646102634681313
        },
        "auroc": {
          "value": 0.6158762360275151,
          "lo": 0.5942953451246389,
          "hi": 0.6370132106595583
        },
        "acc@80": {
          "value": 0.6176470588235294,
          "lo": 0.5971840436530517,
          "hi": 0.638120808284715
        }
      }
    },
    {
      "env": "civilcomments",
      "slice": "all",
      "system": "LFM2.5-VL-1.6B",
      "metrics": {
        "n": {
          "value": 10581.0,
          "lo": 10261.425,
          "hi": 10904.525
        },
        "ce": {
          "value": 0.628772068815973,
          "lo": 0.6216998765989009,
          "hi": 0.6351669416229918
        },
        "kl": {
          "value": 0.397289260046632,
          "lo": 0.39052833898608946,
          "hi": 0.40449370764458126
        },
        "jsd": {
          "value": 0.11740113018175644,
          "lo": 0.11548964447189249,
          "hi": 0.11934102760312093
        },
        "brier": {
          "value": 0.280118354904552,
          "lo": 0.2744577113120764,
          "hi": 0.2855380627483798
        },
        "acc": {
          "value": 0.6832057461487572,
          "lo": 0.6741245818307544,
          "hi": 0.6931629642589503
        },
        "conf": {
          "value": 0.6608092393206801,
          "lo": 0.6586314404096306,
          "hi": 0.6627601171445902
        },
        "ece": {
          "value": 0.044680318138615394,
          "lo": 0.040354284677392396,
          "hi": 0.05704754718799394
        },
        "reliability": {
          "value": 0.00442753772069274,
          "lo": 0.003552214328484263,
          "hi": 0.00812415256967844
        },
        "resolution": {
          "value": 0.004326625220736689,
          "lo": 0.003684353149117342,
          "hi": 0.00702849888079183
        },
        "uncertainty": {
          "value": 0.21643565457807715,
          "lo": 0.21268806824183523,
          "hi": 0.2196806077426754
        },
        "auroc": {
          "value": 0.570141816424234,
          "lo": 0.55657289916221,
          "hi": 0.5825056692924083
        },
        "acc@80": {
          "value": 0.6973419964559953,
          "lo": 0.685354588080431,
          "hi": 0.7075284918032322
        }
      }
    },
    {
      "env": "civilcomments",
      "slice": "all",
      "system": "LFM2.5-VL-1.6B +T",
      "metrics": {
        "n": {
          "value": 10581.0,
          "lo": 10261.425,
          "hi": 10904.525
        },
        "ce": {
          "value": 0.6286493075695474,
          "lo": 0.6216425549183775,
          "hi": 0.6349876756300533
        },
        "kl": {
          "value": 0.39716649880020644,
          "lo": 0.3903972556891716,
          "hi": 0.4043454095770948
        },
        "jsd": {
          "value": 0.11748060993828376,
          "lo": 0.11557479409577277,
          "hi": 0.11941577090821023
        },
        "brier": {
          "value": 0.2801003072251209,
          "lo": 0.27448260909947786,
          "hi": 0.28549944743997047
        },
        "acc": {
          "value": 0.6832057461487572,
          "lo": 0.6741245818307544,
          "hi": 0.6931629642589503
        },
        "conf": {
          "value": 0.6594996163310518,
          "lo": 0.6573364393569618,
          "hi": 0.6614357378208103
        },
        "ece": {
          "value": 0.04538447461741798,
          "lo": 0.040671909795305515,
          "hi": 0.05725331864949602
        },
        "reliability": {
          "value": 0.004429914427853016,
          "lo": 0.003549175662120752,
          "hi": 0.008105769448894466
        },
        "resolution": {
          "value": 0.004326625220736689,
          "lo": 0.003684353149117342,
          "hi": 0.00702849888079183
        },
        "uncertainty": {
          "value": 0.21643565457807715,
          "lo": 0.21268806824183523,
          "hi": 0.2196806077426754
        },
        "auroc": {
          "value": 0.570141816424234,
          "lo": 0.55657289916221,
          "hi": 0.5825056692924083
        },
        "acc@80": {
          "value": 0.6973419964559953,
          "lo": 0.685354588080431,
          "hi": 0.7075284918032322
        }
      }
    },
    {
      "env": "folktables",
      "slice": "all",
      "system": "LFM2.5-VL-1.6B",
      "metrics": {
        "n": {
          "value": 15003.0,
          "lo": 14855.8,
          "hi": 15157.0
        },
        "ce": {
          "value": 0.7461344360201306,
          "lo": 0.7318165582948096,
          "hi": 0.7625133777479692
        },
        "kl": {
          "value": 0.7461344360201306,
          "lo": 0.7318165582948096,
          "hi": 0.7625133777479692
        },
        "jsd": {
          "value": 0.18684880414111565,
          "lo": 0.18358249706605045,
          "hi": 0.1905574729419387
        },
        "brier": {
          "value": 0.5087217043706055,
          "lo": 0.4979805961161123,
          "hi": 0.5205494304915599
        },
        "acc": {
          "value": 0.6536026128107711,
          "lo": 0.6451184668229782,
          "hi": 0.661609588316768
        },
        "conf": {
          "value": 0.8728890623528585,
          "lo": 0.8713695262973349,
          "hi": 0.8741989386944872
        },
        "ece": {
          "value": 0.21928644954208723,
          "lo": 0.2120593261305226,
          "hi": 0.2270807691928449
        },
        "reliability": {
          "value": 0.05965806026922896,
          "lo": 0.05587235775469685,
          "hi": 0.06374055446160587
        },
        "resolution": {
          "value": 0.03126160888216562,
          "lo": 0.029147991406127175,
          "hi": 0.034045515786957116
        },
        "uncertainty": {
          "value": 0.22640623733770432,
          "lo": 0.22388234090058073,
          "hi": 0.2289406268460248
        },
        "auroc": {
          "value": 0.7191105974276959,
          "lo": 0.710611030947418,
          "hi": 0.7276068851737252
        },
        "acc@80": {
          "value": 0.7234627562072988,
          "lo": 0.715027466679644,
          "hi": 0.7316353539303582
        }
      }
    },
    {
      "env": "folktables",
      "slice": "all",
      "system": "LFM2.5-VL-1.6B +T",
      "metrics": {
        "n": {
          "value": 15003.0,
          "lo": 14855.8,
          "hi": 15157.0
        },
        "ce": {
          "value": 0.6121096088424054,
          "lo": 0.605738667381172,
          "hi": 0.6194423479030153
        },
        "kl": {
          "value": 0.6121096088424054,
          "lo": 0.605738667381172,
          "hi": 0.6194423479030153
        },
        "jsd": {
          "value": 0.18423216443856238,
          "lo": 0.182536459553405,
          "hi": 0.1861922559361493
        },
        "brier": {
          "value": 0.42469090692255734,
          "lo": 0.4187951793424778,
          "hi": 0.43151763460653886
        },
        "acc": {
          "value": 0.6536026128107711,
          "lo": 0.6451184668229782,
          "hi": 0.661609588316768
        },
        "conf": {
          "value": 0.7066941692618047,
          "lo": 0.7055402831071527,
          "hi": 0.7076874142613194
        },
        "ece": {
          "value": 0.09931113401963056,
          "lo": 0.0927887134946938,
          "hi": 0.10663302728364965
        },
        "reliability": {
          "value": 0.017519382831588975,
          "lo": 0.015838382170399252,
          "hi": 0.019669253791422578
        },
        "resolution": {
          "value": 0.03126160888216562,
          "lo": 0.029147991406127175,
          "hi": 0.034045515786957116
        },
        "uncertainty": {
          "value": 0.22640623733770432,
          "lo": 0.22388234090058073,
          "hi": 0.2289406268460248
        },
        "auroc": {
          "value": 0.7191105974276959,
          "lo": 0.710611030947418,
          "hi": 0.7276068851737252
        },
        "acc@80": {
          "value": 0.7234627562072988,
          "lo": 0.715027466679644,
          "hi": 0.7316353539303582
        }
      }
    },
    {
      "env": "mcq",
      "slice": "all",
      "system": "LFM2.5-VL-1.6B",
      "metrics": {
        "n": {
          "value": 5064.0,
          "lo": 5064.0,
          "hi": 5064.0
        },
        "ce": {
          "value": 0.572165947776608,
          "lo": 0.5452410135475054,
          "hi": 0.600168726395567
        },
        "kl": {
          "value": 0.572165947776608,
          "lo": 0.5452410135475054,
          "hi": 0.600168726395567
        },
        "jsd": {
          "value": 0.1321212551223715,
          "lo": 0.12725109212655403,
          "hi": 0.137655877383627
        },
        "brier": {
          "value": 0.2907395378003839,
          "lo": 0.277269503680997,
          "hi": 0.3036812445034888
        },
        "acc": {
          "value": 0.7940363349131122,
          "lo": 0.7818868483412322,
          "hi": 0.80451224328594
        },
        "conf": {
          "value": 0.8386109491830801,
          "lo": 0.8331356306114539,
          "hi": 0.8433986392359162
        },
        "ece": {
          "value": 0.04461313826789694,
          "lo": 0.03691930355427648,
          "hi": 0.05483356854248656
        },
        "reliability": {
          "value": 0.0031711189146454073,
          "lo": 0.0023652986228275426,
          "hi": 0.004927669784027484
        },
        "resolution": {
          "value": 0.0415554142512013,
          "lo": 0.037302450812327796,
          "hi": 0.04679214675977566
        },
        "uncertainty": {
          "value": 0.1635426337508641,
          "lo": 0.1572722547911223,
          "hi": 0.1705397950077866
        },
        "auroc": {
          "value": 0.8434959034579483,
          "lo": 0.8314073112444279,
          "hi": 0.8549349683519096
        },
        "acc@80": {
          "value": 0.8765736855097507,
          "lo": 0.8666872377190817,
          "hi": 0.8871883485559121
        }
      }
    },
    {
      "env": "mcq",
      "slice": "all",
      "system": "LFM2.5-VL-1.6B +T",
      "metrics": {
        "n": {
          "value": 5064.0,
          "lo": 5064.0,
          "hi": 5064.0
        },
        "ce": {
          "value": 0.5558825055009654,
          "lo": 0.5348871978845787,
          "hi": 0.5784931928840349
        },
        "kl": {
          "value": 0.5558825055009654,
          "lo": 0.5348871978845787,
          "hi": 0.5784931928840349
        },
        "jsd": {
          "value": 0.14306107981735358,
          "lo": 0.13869103713638892,
          "hi": 0.14781201028160543
        },
        "brier": {
          "value": 0.28702519231409684,
          "lo": 0.275253304876988,
          "hi": 0.2992549336460319
        },
        "acc": {
          "value": 0.7940363349131122,
          "lo": 0.7818868483412322,
          "hi": 0.80451224328594
        },
        "conf": {
          "value": 0.7848499106116366,
          "lo": 0.7784763165245263,
          "hi": 0.7901990883085657
        },
        "ece": {
          "value": 0.013609003904996479,
          "lo": 0.010953061439293286,
          "hi": 0.025846386110621757
        },
        "reliability": {
          "value": 0.0002893597635578177,
          "lo": 0.00021791850356241846,
          "hi": 0.0014217689143205163
        },
        "resolution": {
          "value": 0.041273284636219036,
          "lo": 0.037026098960848956,
          "hi": 0.04646639681337491
        },
        "uncertainty": {
          "value": 0.1635426337508641,
          "lo": 0.1572722547911223,
          "hi": 0.1705397950077866
        },
        "auroc": {
          "value": 0.8431811608422989,
          "lo": 0.8313809563709517,
          "hi": 0.8543887509495475
        },
        "acc@80": {
          "value": 0.8775610960256727,
          "lo": 0.8678042458652184,
          "hi": 0.8874475438163416
        }
      }
    },
    {
      "env": "stsb",
      "slice": "all",
      "system": "LFM2.5-VL-1.6B",
      "metrics": {
        "n": {
          "value": 3676.0,
          "lo": 3150.95,
          "hi": 4793.674999999996
        },
        "ce": {
          "value": 2.3010640888069385,
          "lo": 2.219348265834835,
          "hi": 2.3945236861110115
        },
        "kl": {
          "value": 1.8987602659446827,
          "lo": 1.812395668802827,
          "hi": 1.9966693748004591
        },
        "jsd": {
          "value": 0.3890836419430554,
          "lo": 0.37413400486300785,
          "hi": 0.40635039337373485
        },
        "brier": {
          "value": 0.7514212698668175,
          "lo": 0.7248395195004066,
          "hi": 0.774923581452912
        },
        "acc": {
          "value": 0.23449401523394994,
          "lo": 0.20829456008503736,
          "hi": 0.26186580018791106
        },
        "conf": {
          "value": 0.6277639134356324,
          "lo": 0.610452234221094,
          "hi": 0.6395318968402306
        },
        "ece": {
          "value": 0.3932698982016825,
          "lo": 0.3754240981422377,
          "hi": 0.40815711587016956
        },
        "reliability": {
          "value": 0.16204072167674188,
          "lo": 0.14865970239664933,
          "hi": 0.17464263148459871
        },
        "resolution": {
          "value": 0.015115334255708575,
          "lo": 0.010829255968167148,
          "hi": 0.021201278220300646
        },
        "uncertainty": {
          "value": 0.17950657205341,
          "lo": 0.16490790178615253,
          "hi": 0.19329210120204962
        },
        "auroc": {
          "value": 0.6323157167427694,
          "lo": 0.5873000880206326,
          "hi": 0.6841052866583409
        },
        "acc@80": {
          "value": 0.2862971778306698,
          "lo": 0.2554955400741242,
          "hi": 0.3120790916209867
        }
      }
    },
    {
      "env": "stsb",
      "slice": "all",
      "system": "LFM2.5-VL-1.6B +T",
      "metrics": {
        "n": {
          "value": 3676.0,
          "lo": 3150.95,
          "hi": 4793.674999999996
        },
        "ce": {
          "value": 1.7525020624452756,
          "lo": 1.7332842234632335,
          "hi": 1.7758527468562777
        },
        "kl": {
          "value": 1.35019823958302,
          "lo": 1.3248534405741286,
          "hi": 1.3783798358965833
        },
        "jsd": {
          "value": 0.3586514174344883,
          "lo": 0.3534942812011378,
          "hi": 0.36449631932752286
        },
        "brier": {
          "value": 0.5466196704914243,
          "lo": 0.5345524094586155,
          "hi": 0.5583900186117372
        },
        "acc": {
          "value": 0.23449401523394994,
          "lo": 0.20829456008503736,
          "hi": 0.26186580018791106
        },
        "conf": {
          "value": 0.25416640367558696,
          "lo": 0.2516250643543517,
          "hi": 0.2560133568043584
        },
        "ece": {
          "value": 0.10395879835710332,
          "lo": 0.08579567132893845,
          "hi": 0.1290127644194208
        },
        "reliability": {
          "value": 0.014588727222329737,
          "lo": 0.010642276200674787,
          "hi": 0.021232165150903908
        },
        "resolution": {
          "value": 0.016003579625248947,
          "lo": 0.011654112067711338,
          "hi": 0.022233263244754886
        },
        "uncertainty": {
          "value": 0.17950657205341,
          "lo": 0.16490790178615253,
          "hi": 0.19329210120204962
        },
        "auroc": {
          "value": 0.6143144898642353,
          "lo": 0.5670806578704596,
          "hi": 0.6697986028198386
        },
        "acc@80": {
          "value": 0.28357701462087725,
          "lo": 0.2540760524470549,
          "hi": 0.3115851173964403
        }
      }
    },
    {
      "env": "yelp",
      "slice": "all",
      "system": "LFM2.5-VL-1.6B",
      "metrics": {
        "n": {
          "value": 5000.0,
          "lo": 5000.0,
          "hi": 5000.0
        },
        "ce": {
          "value": 1.2672820743302236,
          "lo": 1.2379575231452973,
          "hi": 1.296280505283855
        },
        "kl": {
          "value": 1.2672820743302236,
          "lo": 1.2379575231452973,
          "hi": 1.296280505283855
        },
        "jsd": {
          "value": 0.3070395576017841,
          "lo": 0.30149149729949076,
          "hi": 0.31204754471832824
        },
        "brier": {
          "value": 0.6588944098752754,
          "lo": 0.6449593526627326,
          "hi": 0.6723762128925471
        },
        "acc": {
          "value": 0.4702,
          "lo": 0.4572,
          "hi": 0.48362
        },
        "conf": {
          "value": 0.6046248717546463,
          "lo": 0.5987460925602913,
          "hi": 0.610359503082037
        },
        "ece": {
          "value": 0.1344248717546463,
          "lo": 0.12321233036130667,
          "hi": 0.14781036010198295
        },
        "reliability": {
          "value": 0.024641956620509477,
          "lo": 0.021457088746067842,
          "hi": 0.029497053853111736
        },
        "resolution": {
          "value": 0.024504319425293557,
          "lo": 0.02105900824547472,
          "hi": 0.02885907580372959
        },
        "uncertainty": {
          "value": 0.24911196000000002,
          "lo": 0.24816815999999997,
          "hi": 0.24973153599999998
        },
        "auroc": {
          "value": 0.663396490477615,
          "lo": 0.6478424101253941,
          "hi": 0.6779021032724923
        },
        "acc@80": {
          "value": 0.51325,
          "lo": 0.49825,
          "hi": 0.52725
        }
      }
    },
    {
      "env": "yelp",
      "slice": "all",
      "system": "LFM2.5-VL-1.6B +T",
      "metrics": {
        "n": {
          "value": 5000.0,
          "lo": 5000.0,
          "hi": 5000.0
        },
        "ce": {
          "value": 1.226441656161317,
          "lo": 1.203748108594829,
          "hi": 1.2478335056564203
        },
        "kl": {
          "value": 1.226441656161317,
          "lo": 1.203748108594829,
          "hi": 1.2478335056564203
        },
        "jsd": {
          "value": 0.3188183609293161,
          "lo": 0.3142419866773621,
          "hi": 0.3229560394339662
        },
        "brier": {
          "value": 0.6366238617464098,
          "lo": 0.6249159437916981,
          "hi": 0.6473524135460489
        },
        "acc": {
          "value": 0.4702,
          "lo": 0.4572,
          "hi": 0.48362
        },
        "conf": {
          "value": 0.5124118862896443,
          "lo": 0.5070885174999723,
          "hi": 0.5181019518532565
        },
        "ece": {
          "value": 0.052707036505233985,
          "lo": 0.04610336615075152,
          "hi": 0.06913588327222285
        },
        "reliability": {
          "value": 0.004453180531319464,
          "lo": 0.0034316601113259246,
          "hi": 0.006909642281025447
        },
        "resolution": {
          "value": 0.024464507157936383,
          "lo": 0.02122354112907011,
          "hi": 0.028876068000622108
        },
        "uncertainty": {
          "value": 0.24911196000000002,
          "lo": 0.24816815999999997,
          "hi": 0.24973153599999998
        },
        "auroc": {
          "value": 0.6644650863009548,
          "lo": 0.6492578389442697,
          "hi": 0.6792196427477908
        },
        "acc@80": {
          "value": 0.513,
          "lo": 0.49836875,
          "hi": 0.52814375
        }
      }
    },
    {
      "env": "boolq",
      "slice": "all",
      "system": "zs-LFM2.5-VL-1.6B-prefill",
      "metrics": {
        "n": {
          "value": 8129.0,
          "lo": 7984.475,
          "hi": 8270.575
        },
        "ce": {
          "value": 0.45809081848121175,
          "lo": 0.4459957462368789,
          "hi": 0.47074164998784135
        },
        "kl": {
          "value": 0.45809081848121175,
          "lo": 0.4459957462368789,
          "hi": 0.47074164998784135
        },
        "jsd": {
          "value": 0.13492659985478037,
          "lo": 0.13182005876327219,
          "hi": 0.13822748077897493
        },
        "brier": {
          "value": 0.29719946416782766,
          "lo": 0.2876802262296285,
          "hi": 0.30709230436429497
        },
        "acc": {
          "value": 0.7844753352195842,
          "lo": 0.7741087364795196,
          "hi": 0.7952461097461098
        },
        "conf": {
          "value": 0.7698005057228845,
          "lo": 0.7658782172520326,
          "hi": 0.7734927521904338
        },
        "ece": {
          "value": 0.017457544466331448,
          "lo": 0.015434384582842886,
          "hi": 0.029130645298296235
        },
        "reliability": {
          "value": 0.00043783593969573645,
          "lo": 0.00035828944809354607,
          "hi": 0.0012241892843678525
        },
        "resolution": {
          "value": 0.020778926965138774,
          "lo": 0.01864961651885824,
          "hi": 0.02375247698322183
        },
        "uncertainty": {
          "value": 0.16907378365170517,
          "lo": 0.16282972658385433,
          "hi": 0.17486439983689395
        },
        "auroc": {
          "value": 0.7424820344660427,
          "lo": 0.7301943723927901,
          "hi": 0.7563721693720247
        },
        "acc@80": {
          "value": 0.8422266646163309,
          "lo": 0.8309900400093337,
          "hi": 0.8519479967797275
        }
      }
    },
    {
      "env": "boolq",
      "slice": "all",
      "system": "zs-LFM2.5-VL-1.6B-prefill +T",
      "metrics": {
        "n": {
          "value": 8129.0,
          "lo": 7984.475,
          "hi": 8270.575
        },
        "ce": {
          "value": 0.4574282843520093,
          "lo": 0.44385529922606226,
          "hi": 0.47141889239347284
        },
        "kl": {
          "value": 0.4574282843520093,
          "lo": 0.44385529922606226,
          "hi": 0.47141889239347284
        },
        "jsd": {
          "value": 0.13152558032176487,
          "lo": 0.128123333032652,
          "hi": 0.13509654541911817
        },
        "brier": {
          "value": 0.2966681446394566,
          "lo": 0.286429025745625,
          "hi": 0.3072686901546995
        },
        "acc": {
          "value": 0.7844753352195842,
          "lo": 0.7741087364795196,
          "hi": 0.7952461097461098
        },
        "conf": {
          "value": 0.7898864226759795,
          "lo": 0.7857718409246228,
          "hi": 0.7936401831584182
        },
        "ece": {
          "value": 0.010848552134439696,
          "lo": 0.010703686129022155,
          "hi": 0.02327767605806169
        },
        "reliability": {
          "value": 0.00017561707931481276,
          "lo": 0.00019444645099059272,
          "hi": 0.0008448418020550343
        },
        "resolution": {
          "value": 0.020778926965138774,
          "lo": 0.01864961651885824,
          "hi": 0.02375247698322183
        },
        "uncertainty": {
          "value": 0.16907378365170517,
          "lo": 0.16282972658385433,
          "hi": 0.17486439983689395
        },
        "auroc": {
          "value": 0.7424820344660427,
          "lo": 0.7301943723927901,
          "hi": 0.7563721693720247
        },
        "acc@80": {
          "value": 0.8422266646163309,
          "lo": 0.8309900400093337,
          "hi": 0.8519479967797275
        }
      }
    },
    {
      "env": "chaosnli",
      "slice": "all",
      "system": "zs-LFM2.5-VL-1.6B-prefill",
      "metrics": {
        "n": {
          "value": 2763.0,
          "lo": 2732.475,
          "hi": 2798.525
        },
        "ce": {
          "value": 0.9518350511907218,
          "lo": 0.9283674771218429,
          "hi": 0.9747391947504885
        },
        "kl": {
          "value": 0.4214749811214598,
          "lo": 0.4022965505307716,
          "hi": 0.4429000525633131
        },
        "jsd": {
          "value": 0.09261838684960924,
          "lo": 0.08886798776282236,
          "hi": 0.09671742531030467
        },
        "brier": {
          "value": 0.25716765006067627,
          "lo": 0.2452494254966154,
          "hi": 0.26993159240426684
        },
        "acc": {
          "value": 0.6058631921824105,
          "lo": 0.5882627514525123,
          "hi": 0.623616533456431
        },
        "conf": {
          "value": 0.7285714013353017,
          "lo": 0.7234689812577547,
          "hi": 0.7342638359808698
        },
        "ece": {
          "value": 0.12270820915289131,
          "lo": 0.10870676878938139,
          "hi": 0.14057291935152566
        },
        "reliability": {
          "value": 0.018860741503653168,
          "lo": 0.0153226981630458,
          "hi": 0.02550495911405989
        },
        "resolution": {
          "value": 0.013356532305657276,
          "lo": 0.010626462844700021,
          "hi": 0.018205897395722793
        },
        "uncertainty": {
          "value": 0.23879298454095002,
          "lo": 0.234718782714523,
          "hi": 0.242209681238806
        },
        "auroc": {
          "value": 0.6302217899643772,
          "lo": 0.6092174144235512,
          "hi": 0.6487577196468025
        },
        "acc@80": {
          "value": 0.644343891402715,
          "lo": 0.6252236392170022,
          "hi": 0.6633401205261578
        }
      }
    },
    {
      "env": "chaosnli",
      "slice": "all",
      "system": "zs-LFM2.5-VL-1.6B-prefill +T",
      "metrics": {
        "n": {
          "value": 2763.0,
          "lo": 2732.475,
          "hi": 2798.525
        },
        "ce": {
          "value": 0.8347258254601914,
          "lo": 0.8226522855169598,
          "hi": 0.8456516996319347
        },
        "kl": {
          "value": 0.3043657553909295,
          "lo": 0.29452635459737536,
          "hi": 0.31487484659061205
        },
        "jsd": {
          "value": 0.08303659510340684,
          "lo": 0.08057666355029988,
          "hi": 0.08582469297651862
        },
        "brier": {
          "value": 0.2044370114518093,
          "lo": 0.19698183969822086,
          "hi": 0.21250505908078474
        },
        "acc": {
          "value": 0.6058631921824105,
          "lo": 0.5882627514525123,
          "hi": 0.623616533456431
        },
        "conf": {
          "value": 0.5782082809309398,
          "lo": 0.5741746859961775,
          "hi": 0.5825492156087732
        },
        "ece": {
          "value": 0.03619162070865312,
          "lo": 0.03401204560331832,
          "hi": 0.06117888941289102
        },
        "reliability": {
          "value": 0.002101243035652229,
          "lo": 0.001923223605356818,
          "hi": 0.005806942654502659
        },
        "resolution": {
          "value": 0.012847809609936024,
          "lo": 0.010548778386390832,
          "hi": 0.01850531944868528
        },
        "uncertainty": {
          "value": 0.23879298454095002,
          "lo": 0.234718782714523,
          "hi": 0.242209681238806
        },
        "auroc": {
          "value": 0.630666390197182,
          "lo": 0.6093681868939779,
          "hi": 0.6500422528507463
        },
        "acc@80": {
          "value": 0.6475113122171946,
          "lo": 0.6277296313785591,
          "hi": 0.6668965468301858
        }
      }
    },
    {
      "env": "civilcomments",
      "slice": "all",
      "system": "zs-LFM2.5-VL-1.6B-prefill",
      "metrics": {
        "n": {
          "value": 10581.0,
          "lo": 10261.425,
          "hi": 10904.525
        },
        "ce": {
          "value": 0.6502473360734663,
          "lo": 0.6447503319703678,
          "hi": 0.6555783078692586
        },
        "kl": {
          "value": 0.41876452730412533,
          "lo": 0.41192029397300456,
          "hi": 0.42577051054253134
        },
        "jsd": {
          "value": 0.12545754668024597,
          "lo": 0.12342769463728959,
          "hi": 0.12751187936910485
        },
        "brier": {
          "value": 0.3021940400965815,
          "lo": 0.29682435114805916,
          "hi": 0.30791112680685717
        },
        "acc": {
          "value": 0.6406766846233816,
          "lo": 0.6309652583155962,
          "hi": 0.6504697018498695
        },
        "conf": {
          "value": 0.6305742715462069,
          "lo": 0.6286774044784078,
          "hi": 0.6323416374505513
        },
        "ece": {
          "value": 0.03474427794995554,
          "lo": 0.03138912308177586,
          "hi": 0.04557950196961081
        },
        "reliability": {
          "value": 0.0026833209329026487,
          "lo": 0.002160425696622576,
          "hi": 0.003972453846252947
        },
        "resolution": {
          "value": 0.0033928532405596472,
          "lo": 0.002752364853399364,
          "hi": 0.005004430122386786
        },
        "uncertainty": {
          "value": 0.23021007040337363,
          "lo": 0.22735886827276156,
          "hi": 0.23284807778724742
        },
        "auroc": {
          "value": 0.5576777550250918,
          "lo": 0.5472387826952736,
          "hi": 0.5695659157795933
        },
        "acc@80": {
          "value": 0.6537507383343177,
          "lo": 0.6412779888193662,
          "hi": 0.6635854397027292
        }
      }
    },
    {
      "env": "civilcomments",
      "slice": "all",
      "system": "zs-LFM2.5-VL-1.6B-prefill +T",
      "metrics": {
        "n": {
          "value": 10581.0,
          "lo": 10261.425,
          "hi": 10904.525
        },
        "ce": {
          "value": 0.6502288177519915,
          "lo": 0.6447428064421498,
          "hi": 0.6555474234689279
        },
        "kl": {
          "value": 0.4187460089826506,
          "lo": 0.41190585434423493,
          "hi": 0.4257500233279973
        },
        "jsd": {
          "value": 0.12546885045847098,
          "lo": 0.1234402570863187,
          "hi": 0.1275238311101115
        },
        "brier": {
          "value": 0.3021844160960673,
          "lo": 0.2968225989617081,
          "hi": 0.3078932463327301
        },
        "acc": {
          "value": 0.6406766846233816,
          "lo": 0.6309652583155962,
          "hi": 0.6504697018498695
        },
        "conf": {
          "value": 0.6303210618646045,
          "lo": 0.6284271536306199,
          "hi": 0.6320855428907367
        },
        "ece": {
          "value": 0.034696598410399876,
          "lo": 0.031366328712397716,
          "hi": 0.04553824327196812
        },
        "reliability": {
          "value": 0.0026774160508352897,
          "lo": 0.002155129499152936,
          "hi": 0.0039658681084932335
        },
        "resolution": {
          "value": 0.0033928532405596472,
          "lo": 0.002752364853399364,
          "hi": 0.005004430122386786
        },
        "uncertainty": {
          "value": 0.23021007040337363,
          "lo": 0.22735886827276156,
          "hi": 0.23284807778724742
        },
        "auroc": {
          "value": 0.5576777550250918,
          "lo": 0.5472387826952736,
          "hi": 0.5695659157795933
        },
        "acc@80": {
          "value": 0.6537507383343177,
          "lo": 0.6412779888193662,
          "hi": 0.6635854397027292
        }
      }
    },
    {
      "env": "folktables",
      "slice": "all",
      "system": "zs-LFM2.5-VL-1.6B-prefill",
      "metrics": {
        "n": {
          "value": 15003.0,
          "lo": 14855.8,
          "hi": 15157.0
        },
        "ce": {
          "value": 0.8197066959026252,
          "lo": 0.8030192075311231,
          "hi": 0.8382327868642858
        },
        "kl": {
          "value": 0.8197066959026252,
          "lo": 0.8030192075311231,
          "hi": 0.8382327868642858
        },
        "jsd": {
          "value": 0.1938454187339975,
          "lo": 0.1902161653193078,
          "hi": 0.19782609427049014
        },
        "brier": {
          "value": 0.5447034447983794,
          "lo": 0.533082550576082,
          "hi": 0.5573901888815646
        },
        "acc": {
          "value": 0.6536026128107711,
          "lo": 0.6451184668229782,
          "hi": 0.661609588316768
        },
        "conf": {
          "value": 0.9027565910737848,
          "lo": 0.901633137398959,
          "hi": 0.9037675302880445
        },
        "ece": {
          "value": 0.24915397826301364,
          "lo": 0.2419406006958448,
          "hi": 0.25708588876344846
        },
        "reliability": {
          "value": 0.07734302902894433,
          "lo": 0.07311883088930139,
          "hi": 0.08198136422133916
        },
        "resolution": {
          "value": 0.031099067312709942,
          "lo": 0.028930295369946016,
          "hi": 0.033717553045725074
        },
        "uncertainty": {
          "value": 0.22640623733770432,
          "lo": 0.22388234090058073,
          "hi": 0.2289406268460248
        },
        "auroc": {
          "value": 0.7213528267123783,
          "lo": 0.7130525215248933,
          "hi": 0.7297975155813541
        },
        "acc@80": {
          "value": 0.7195467422096318,
          "lo": 0.7107755151525855,
          "hi": 0.7283890372652869
        }
      }
    },
    {
      "env": "folktables",
      "slice": "all",
      "system": "zs-LFM2.5-VL-1.6B-prefill +T",
      "metrics": {
        "n": {
          "value": 15003.0,
          "lo": 14855.8,
          "hi": 15157.0
        },
        "ce": {
          "value": 0.6167797650432055,
          "lo": 0.6105579967897169,
          "hi": 0.6238108651779094
        },
        "kl": {
          "value": 0.6167797650432055,
          "lo": 0.6105579967897169,
          "hi": 0.6238108651779094
        },
        "jsd": {
          "value": 0.1860303969380419,
          "lo": 0.18435871097790205,
          "hi": 0.1879150410105986
        },
        "brier": {
          "value": 0.4287724273270234,
          "lo": 0.42292768984183027,
          "hi": 0.4353708842047484
        },
        "acc": {
          "value": 0.6536026128107711,
          "lo": 0.6451184668229782,
          "hi": 0.661609588316768
        },
        "conf": {
          "value": 0.7026956371286125,
          "lo": 0.7017580592058847,
          "hi": 0.7035779274954069
        },
        "ece": {
          "value": 0.11113349067205512,
          "lo": 0.10483935635734597,
          "hi": 0.11866877782849507
        },
        "reliability": {
          "value": 0.019291929561390565,
          "lo": 0.01740403521074185,
          "hi": 0.021570290379250588
        },
        "resolution": {
          "value": 0.031099067312709942,
          "lo": 0.028930295369946016,
          "hi": 0.033717553045725074
        },
        "uncertainty": {
          "value": 0.22640623733770432,
          "lo": 0.22388234090058073,
          "hi": 0.2289406268460248
        },
        "auroc": {
          "value": 0.7213528267123783,
          "lo": 0.7130525215248933,
          "hi": 0.7297975155813541
        },
        "acc@80": {
          "value": 0.7195467422096318,
          "lo": 0.7107755151525855,
          "hi": 0.7283890372652869
        }
      }
    },
    {
      "env": "mcq",
      "slice": "all",
      "system": "zs-LFM2.5-VL-1.6B-prefill",
      "metrics": {
        "n": {
          "value": 5064.0,
          "lo": 5064.0,
          "hi": 5064.0
        },
        "ce": {
          "value": 0.567471638295474,
          "lo": 0.5355246464459119,
          "hi": 0.5998490333572554
        },
        "kl": {
          "value": 0.567471638295474,
          "lo": 0.5355246464459119,
          "hi": 0.5998490333572554
        },
        "jsd": {
          "value": 0.12043067454541828,
          "lo": 0.11526442191156205,
          "hi": 0.12609448505334858
        },
        "brier": {
          "value": 0.27981321356917266,
          "lo": 0.26538361838399294,
          "hi": 0.2954207827576226
        },
        "acc": {
          "value": 0.8104265402843602,
          "lo": 0.8001579778830964,
          "hi": 0.8204038309636651
        },
        "conf": {
          "value": 0.8767070750842727,
          "lo": 0.8721753720947398,
          "hi": 0.8810695888627598
        },
        "ece": {
          "value": 0.0663454392910192,
          "lo": 0.05677600162439224,
          "hi": 0.0766042745597131
        },
        "reliability": {
          "value": 0.007613211206775605,
          "lo": 0.005713414187734919,
          "hi": 0.01029174037611074
        },
        "resolution": {
          "value": 0.03825688826641818,
          "lo": 0.034504085352924735,
          "hi": 0.04331040736302365
        },
        "uncertainty": {
          "value": 0.15363536308708248,
          "lo": 0.14734137537934652,
          "hi": 0.1599051883131306
        },
        "auroc": {
          "value": 0.8459082094704353,
          "lo": 0.8333766358908723,
          "hi": 0.8573552144558576
        },
        "acc@80": {
          "value": 0.8908911379906196,
          "lo": 0.8797827696864972,
          "hi": 0.8996605776351518
        }
      }
    },
    {
      "env": "mcq",
      "slice": "all",
      "system": "zs-LFM2.5-VL-1.6B-prefill +T",
      "metrics": {
        "n": {
          "value": 5064.0,
          "lo": 5064.0,
          "hi": 5064.0
        },
        "ce": {
          "value": 0.5269270169220264,
          "lo": 0.5048392107778062,
          "hi": 0.5501583389985965
        },
        "kl": {
          "value": 0.5269270169220264,
          "lo": 0.5048392107778062,
          "hi": 0.5501583389985965
        },
        "jsd": {
          "value": 0.13504835666232706,
          "lo": 0.13062475564306758,
          "hi": 0.1398825440824005
        },
        "brier": {
          "value": 0.2693563783138094,
          "lo": 0.2577888016783025,
          "hi": 0.28193399649932643
        },
        "acc": {
          "value": 0.8104265402843602,
          "lo": 0.8001579778830964,
          "hi": 0.8204038309636651
        },
        "conf": {
          "value": 0.8002951632916763,
          "lo": 0.7950182083441955,
          "hi": 0.8050914600543674
        },
        "ece": {
          "value": 0.020351312816510297,
          "lo": 0.01712466743891839,
          "hi": 0.030929943004074326
        },
        "reliability": {
          "value": 0.00048451933467936613,
          "lo": 0.0004253497776299618,
          "hi": 0.0013789222022809087
        },
        "resolution": {
          "value": 0.03855886462790291,
          "lo": 0.03457260919057467,
          "hi": 0.043692728924570434
        },
        "uncertainty": {
          "value": 0.15363536308708248,
          "lo": 0.14734137537934652,
          "hi": 0.1599051883131306
        },
        "auroc": {
          "value": 0.8465923743096166,
          "lo": 0.8337513744168854,
          "hi": 0.8583663363199558
        },
        "acc@80": {
          "value": 0.8913848432485806,
          "lo": 0.8813811404591458,
          "hi": 0.9008948407800543
        }
      }
    },
    {
      "env": "stsb",
      "slice": "all",
      "system": "zs-LFM2.5-VL-1.6B-prefill",
      "metrics": {
        "n": {
          "value": 3676.0,
          "lo": 3150.95,
          "hi": 4793.674999999996
        },
        "ce": {
          "value": 2.4949856820898892,
          "lo": 2.403966048289237,
          "hi": 2.5998394086267878
        },
        "kl": {
          "value": 2.092681859227634,
          "lo": 1.997292570579143,
          "hi": 2.201441225034531
        },
        "jsd": {
          "value": 0.3968087037540929,
          "lo": 0.38052443084813153,
          "hi": 0.4154332315379364
        },
        "brier": {
          "value": 0.824537909273267,
          "lo": 0.794885561707944,
          "hi": 0.850379683918623
        },
        "acc": {
          "value": 0.23449401523394994,
          "lo": 0.20829456008503736,
          "hi": 0.26186580018791106
        },
        "conf": {
          "value": 0.7117134036984745,
          "lo": 0.6963201652904776,
          "hi": 0.7221477089054564
        },
        "ece": {
          "value": 0.4772193884645245,
          "lo": 0.4590508668907983,
          "hi": 0.4923052034116568
        },
        "reliability": {
          "value": 0.23199332828556152,
          "lo": 0.21560957449199486,
          "hi": 0.24716219570305237
        },
        "resolution": {
          "value": 0.015095316428784285,
          "lo": 0.01132341751077712,
          "hi": 0.020081745389184037
        },
        "uncertainty": {
          "value": 0.17950657205341,
          "lo": 0.16490790178615253,
          "hi": 0.19329210120204962
        },
        "auroc": {
          "value": 0.6818591827076088,
          "lo": 0.6393946118136222,
          "hi": 0.729803709649928
        },
        "acc@80": {
          "value": 0.286977218633118,
          "lo": 0.2565269320815032,
          "hi": 0.3127481120442128
        }
      }
    },
    {
      "env": "stsb",
      "slice": "all",
      "system": "zs-LFM2.5-VL-1.6B-prefill +T",
      "metrics": {
        "n": {
          "value": 3676.0,
          "lo": 3150.95,
          "hi": 4793.674999999996
        },
        "ce": {
          "value": 1.749101240806852,
          "lo": 1.7283041885675015,
          "hi": 1.774187438110359
        },
        "kl": {
          "value": 1.3467974179445963,
          "lo": 1.3195490845977145,
          "hi": 1.3764656306697425
        },
        "jsd": {
          "value": 0.3572562508931146,
          "lo": 0.3517192084644762,
          "hi": 0.36339969260996136
        },
        "brier": {
          "value": 0.5444155979698277,
          "lo": 0.5319484618563186,
          "hi": 0.5569032296162743
        },
        "acc": {
          "value": 0.23449401523394994,
          "lo": 0.20829456008503736,
          "hi": 0.26186580018791106
        },
        "conf": {
          "value": 0.26412474790906104,
          "lo": 0.26161907123082884,
          "hi": 0.2658817355503822
        },
        "ece": {
          "value": 0.09120896808132307,
          "lo": 0.07239989869480402,
          "hi": 0.11569162818968533
        },
        "reliability": {
          "value": 0.013670878123295767,
          "lo": 0.009786623552048445,
          "hi": 0.01985571136586088
        },
        "resolution": {
          "value": 0.01538556123039212,
          "lo": 0.011344999347197309,
          "hi": 0.02065854047173325
        },
        "uncertainty": {
          "value": 0.17950657205341,
          "lo": 0.16490790178615253,
          "hi": 0.19329210120204962
        },
        "auroc": {
          "value": 0.6692634771122841,
          "lo": 0.6256132861111295,
          "hi": 0.7196155209077784
        },
        "acc@80": {
          "value": 0.2852771166269976,
          "lo": 0.2552077935193685,
          "hi": 0.31323414252153486
        }
      }
    },
    {
      "env": "yelp",
      "slice": "all",
      "system": "zs-LFM2.5-VL-1.6B-prefill",
      "metrics": {
        "n": {
          "value": 5000.0,
          "lo": 5000.0,
          "hi": 5000.0
        },
        "ce": {
          "value": 1.2795574641538,
          "lo": 1.2517782952036682,
          "hi": 1.3058245794645162
        },
        "kl": {
          "value": 1.2795574641538,
          "lo": 1.2517782952036682,
          "hi": 1.3058245794645162
        },
        "jsd": {
          "value": 0.31409994204249014,
          "lo": 0.3085733024183901,
          "hi": 0.3189092146477358
        },
        "brier": {
          "value": 0.6701537111869355,
          "lo": 0.6550587774857013,
          "hi": 0.6837073153760602
        },
        "acc": {
          "value": 0.4778,
          "lo": 0.463495,
          "hi": 0.4916
        },
        "conf": {
          "value": 0.6009585103601217,
          "lo": 0.5957442806113511,
          "hi": 0.6065007946558297
        },
        "ece": {
          "value": 0.13250732738077642,
          "lo": 0.12254822125673294,
          "hi": 0.14733026904314755
        },
        "reliability": {
          "value": 0.02536507914825806,
          "lo": 0.021953854325576682,
          "hi": 0.030694678524850767
        },
        "resolution": {
          "value": 0.019443062968837098,
          "lo": 0.01639166898892588,
          "hi": 0.02351216928237807
        },
        "uncertainty": {
          "value": 0.24950716,
          "lo": 0.248667375,
          "hi": 0.24992943999999997
        },
        "auroc": {
          "value": 0.6292576132885325,
          "lo": 0.6118137052534594,
          "hi": 0.6438647698650272
        },
        "acc@80": {
          "value": 0.504,
          "lo": 0.48886875,
          "hi": 0.5198812500000001
        }
      }
    },
    {
      "env": "yelp",
      "slice": "all",
      "system": "zs-LFM2.5-VL-1.6B-prefill +T",
      "metrics": {
        "n": {
          "value": 5000.0,
          "lo": 5000.0,
          "hi": 5000.0
        },
        "ce": {
          "value": 1.2467077560785884,
          "lo": 1.2253371183672128,
          "hi": 1.2665288437688031
        },
        "kl": {
          "value": 1.2467077560785884,
          "lo": 1.2253371183672128,
          "hi": 1.2665288437688031
        },
        "jsd": {
          "value": 0.3246998353081678,
          "lo": 0.3201356534839991,
          "hi": 0.3287172198525013
        },
        "brier": {
          "value": 0.6477096486623636,
          "lo": 0.6355266875009784,
          "hi": 0.6583932973272311
        },
        "acc": {
          "value": 0.4778,
          "lo": 0.463495,
          "hi": 0.4916
        },
        "conf": {
          "value": 0.512566381017171,
          "lo": 0.5078738480498601,
          "hi": 0.5175937098496847
        },
        "ece": {
          "value": 0.07436672394824685,
          "lo": 0.06416638063069739,
          "hi": 0.09020058880391907
        },
        "reliability": {
          "value": 0.006848674886070009,
          "lo": 0.005376357899329489,
          "hi": 0.009904528379284263
        },
        "resolution": {
          "value": 0.019817199380698064,
          "lo": 0.016920157765347506,
          "hi": 0.02387091143686241
        },
        "uncertainty": {
          "value": 0.24950716,
          "lo": 0.248667375,
          "hi": 0.24992943999999997
        },
        "auroc": {
          "value": 0.6293416990518429,
          "lo": 0.6121168328862725,
          "hi": 0.6439684894481816
        },
        "acc@80": {
          "value": 0.50325,
          "lo": 0.48861875,
          "hi": 0.5188812500000001
        }
      }
    },
    {
      "env": "boolq",
      "slice": "all",
      "system": "LFM2.5-VL-3B",
      "metrics": {
        "n": {
          "value": 8129.0,
          "lo": 7984.475,
          "hi": 8270.575
        },
        "ce": {
          "value": 0.5328583476623585,
          "lo": 0.5034779386975547,
          "hi": 0.5645709102211561
        },
        "kl": {
          "value": 0.5328583476623585,
          "lo": 0.5034779386975547,
          "hi": 0.5645709102211561
        },
        "jsd": {
          "value": 0.10291671049966716,
          "lo": 0.09828475981022308,
          "hi": 0.10777257902109055
        },
        "brier": {
          "value": 0.2784675294234334,
          "lo": 0.26475431357101853,
          "hi": 0.2925826290127473
        },
        "acc": {
          "value": 0.8294993234100135,
          "lo": 0.8200077032444733,
          "hi": 0.8384865413882864
        },
        "conf": {
          "value": 0.9367556756295641,
          "lo": 0.9339936626756461,
          "hi": 0.939232940598248
        },
        "ece": {
          "value": 0.10725635221955052,
          "lo": 0.0988807214514619,
          "hi": 0.11561362088654317
        },
        "reliability": {
          "value": 0.016289727611228377,
          "lo": 0.014055659763456274,
          "hi": 0.01910470563320159
        },
        "resolution": {
          "value": 0.018139487624438805,
          "lo": 0.016009183270894112,
          "hi": 0.020861142138065423
        },
        "uncertainty": {
          "value": 0.14143019587234332,
          "lo": 0.1354268563237357,
          "hi": 0.14759506919989493
        },
        "auroc": {
          "value": 0.7673957322852474,
          "lo": 0.754213355432779,
          "hi": 0.7805694196890747
        },
        "acc@80": {
          "value": 0.8829770874980778,
          "lo": 0.8731227133062054,
          "hi": 0.8917302633474256
        }
      }
    },
    {
      "env": "boolq",
      "slice": "all",
      "system": "LFM2.5-VL-3B +T",
      "metrics": {
        "n": {
          "value": 8129.0,
          "lo": 7984.475,
          "hi": 8270.575
        },
        "ce": {
          "value": 0.39341402832620836,
          "lo": 0.3793506116651023,
          "hi": 0.40836095134841677
        },
        "kl": {
          "value": 0.39341402832620836,
          "lo": 0.3793506116651023,
          "hi": 0.40836095134841677
        },
        "jsd": {
          "value": 0.1096984527809455,
          "lo": 0.10643592813274551,
          "hi": 0.11335693290663253
        },
        "brier": {
          "value": 0.24814633560253155,
          "lo": 0.2377302397014397,
          "hi": 0.25936932864393564
        },
        "acc": {
          "value": 0.8294993234100135,
          "lo": 0.8200077032444733,
          "hi": 0.8384865413882864
        },
        "conf": {
          "value": 0.8452801117521198,
          "lo": 0.8419216373999594,
          "hi": 0.8483471561336406
        },
        "ece": {
          "value": 0.02446798654186068,
          "lo": 0.020447147595451493,
          "hi": 0.03533475737885574
        },
        "reliability": {
          "value": 0.0011106918261193821,
          "lo": 0.0007329247301666026,
          "hi": 0.002098926828915632
        },
        "resolution": {
          "value": 0.018139487624438805,
          "lo": 0.016009183270894112,
          "hi": 0.020861142138065423
        },
        "uncertainty": {
          "value": 0.14143019587234332,
          "lo": 0.1354268563237357,
          "hi": 0.14759506919989493
        },
        "auroc": {
          "value": 0.7673957322852474,
          "lo": 0.754213355432779,
          "hi": 0.7805694196890747
        },
        "acc@80": {
          "value": 0.8829770874980778,
          "lo": 0.8731227133062054,
          "hi": 0.8917302633474256
        }
      }
    },
    {
      "env": "chaosnli",
      "slice": "all",
      "system": "LFM2.5-VL-3B",
      "metrics": {
        "n": {
          "value": 2763.0,
          "lo": 2732.475,
          "hi": 2798.525
        },
        "ce": {
          "value": 1.4463869206613593,
          "lo": 1.3969440713345973,
          "hi": 1.496035654471535
        },
        "kl": {
          "value": 0.9160268505920974,
          "lo": 0.8695692282664632,
          "hi": 0.9636039120549132
        },
        "jsd": {
          "value": 0.12339426634731591,
          "lo": 0.11785326462702327,
          "hi": 0.1285760716912485
        },
        "brier": {
          "value": 0.3256343876098984,
          "lo": 0.30917852060517886,
          "hi": 0.3416501268543098
        },
        "acc": {
          "value": 0.66847629388346,
          "lo": 0.6527744206804789,
          "hi": 0.685302583509242
        },
        "conf": {
          "value": 0.8955592298977969,
          "lo": 0.8907141895012433,
          "hi": 0.901283667606361
        },
        "ece": {
          "value": 0.22708293601433696,
          "lo": 0.21198514666343163,
          "hi": 0.2429271002362792
        },
        "reliability": {
          "value": 0.057846081841530395,
          "lo": 0.05095878852089352,
          "hi": 0.06702085714378257
        },
        "resolution": {
          "value": 0.01831117311529517,
          "lo": 0.015418712607187562,
          "hi": 0.02377916072160715
        },
        "uncertainty": {
          "value": 0.22161573839929402,
          "lo": 0.21566295244340217,
          "hi": 0.22665997607375962
        },
        "auroc": {
          "value": 0.6697400245411537,
          "lo": 0.649998624891882,
          "hi": 0.6903150321308094
        },
        "acc@80": {
          "value": 0.716289592760181,
          "lo": 0.6996120589974848,
          "hi": 0.7357207615593835
        }
      }
    },
    {
      "env": "chaosnli",
      "slice": "all",
      "system": "LFM2.5-VL-3B +T",
      "metrics": {
        "n": {
          "value": 2763.0,
          "lo": 2732.475,
          "hi": 2798.525
        },
        "ce": {
          "value": 0.7944889838081668,
          "lo": 0.7817771260528582,
          "hi": 0.8063845883673213
        },
        "kl": {
          "value": 0.26412891373890496,
          "lo": 0.2543928784905818,
          "hi": 0.274557945037211
        },
        "jsd": {
          "value": 0.07146401553311572,
          "lo": 0.06888100386968574,
          "hi": 0.07407442993902372
        },
        "brier": {
          "value": 0.1728068690748953,
          "lo": 0.16526219560988412,
          "hi": 0.1810891866882665
        },
        "acc": {
          "value": 0.66847629388346,
          "lo": 0.6527744206804789,
          "hi": 0.685302583509242
        },
        "conf": {
          "value": 0.6224814034196495,
          "lo": 0.618021338576969,
          "hi": 0.6276350420443061
        },
        "ece": {
          "value": 0.04958543734846641,
          "lo": 0.04367292327058699,
          "hi": 0.0705175026939239
        },
        "reliability": {
          "value": 0.0036959336504425217,
          "lo": 0.003028344555863666,
          "hi": 0.0073011007642513155
        },
        "resolution": {
          "value": 0.0162851595288571,
          "lo": 0.013386979538948246,
          "hi": 0.02120530567234291
        },
        "uncertainty": {
          "value": 0.22161573839929402,
          "lo": 0.21566295244340217,
          "hi": 0.22665997607375962
        },
        "auroc": {
          "value": 0.6575504240323622,
          "lo": 0.6363788080021665,
          "hi": 0.6796131873409706
        },
        "acc@80": {
          "value": 0.7180995475113122,
          "lo": 0.6998164435534286,
          "hi": 0.7346567620671158
        }
      }
    },
    {
      "env": "civilcomments",
      "slice": "all",
      "system": "LFM2.5-VL-3B",
      "metrics": {
        "n": {
          "value": 10581.0,
          "lo": 10261.425,
          "hi": 10904.525
        },
        "ce": {
          "value": 0.5077022501274139,
          "lo": 0.4961021782471906,
          "hi": 0.5180279389922728
        },
        "kl": {
          "value": 0.27621944135807286,
          "lo": 0.26730500664356843,
          "hi": 0.2835696862919652
        },
        "jsd": {
          "value": 0.07257758135861772,
          "lo": 0.07055121962272641,
          "hi": 0.07418253016567902
        },
        "brier": {
          "value": 0.17210397486649306,
          "lo": 0.16552973030690482,
          "hi": 0.17740769611117288
        },
        "acc": {
          "value": 0.8038937718552122,
          "lo": 0.795871605523184,
          "hi": 0.8129462172944398
        },
        "conf": {
          "value": 0.8094657350832546,
          "lo": 0.8063933901096233,
          "hi": 0.8122725926216197
        },
        "ece": {
          "value": 0.02503634923212628,
          "lo": 0.023475550226927357,
          "hi": 0.03463331161127155
        },
        "reliability": {
          "value": 0.0014436016102529808,
          "lo": 0.0010643736014497817,
          "hi": 0.002367815593946149
        },
        "resolution": {
          "value": 0.012833172824697326,
          "lo": 0.011369842633845467,
          "hi": 0.014824249058854993
        },
        "uncertainty": {
          "value": 0.15764857542761226,
          "lo": 0.15206466487253756,
          "hi": 0.1624599924504572
        },
        "auroc": {
          "value": 0.6998776200499152,
          "lo": 0.6875208159261627,
          "hi": 0.7120482908431496
        },
        "acc@80": {
          "value": 0.8473715298287064,
          "lo": 0.8399738848348878,
          "hi": 0.8569233002961383
        }
      }
    },
    {
      "env": "civilcomments",
      "slice": "all",
      "system": "LFM2.5-VL-3B +T",
      "metrics": {
        "n": {
          "value": 10581.0,
          "lo": 10261.425,
          "hi": 10904.525
        },
        "ce": {
          "value": 0.49676150002708036,
          "lo": 0.4872737952681122,
          "hi": 0.5053113122868408
        },
        "kl": {
          "value": 0.2652786912577394,
          "lo": 0.25799358231107855,
          "hi": 0.2712157832203521
        },
        "jsd": {
          "value": 0.07525692086753505,
          "lo": 0.07339140236779222,
          "hi": 0.07673566141619703
        },
        "brier": {
          "value": 0.16868331824137972,
          "lo": 0.1629474417311238,
          "hi": 0.17327347459413364
        },
        "acc": {
          "value": 0.8038937718552122,
          "lo": 0.795871605523184,
          "hi": 0.8129462172944398
        },
        "conf": {
          "value": 0.7700251446366264,
          "lo": 0.767086855986615,
          "hi": 0.7728444308829732
        },
        "ece": {
          "value": 0.035339698457243184,
          "lo": 0.030404589401963122,
          "hi": 0.044576461852605574
        },
        "reliability": {
          "value": 0.002158982108121902,
          "lo": 0.0015709597177534181,
          "hi": 0.003239906426687996
        },
        "resolution": {
          "value": 0.012833172824697326,
          "lo": 0.011369842633845467,
          "hi": 0.014824249058854993
        },
        "uncertainty": {
          "value": 0.15764857542761226,
          "lo": 0.15206466487253756,
          "hi": 0.1624599924504572
        },
        "auroc": {
          "value": 0.6998776200499152,
          "lo": 0.6875208159261627,
          "hi": 0.7120482908431496
        },
        "acc@80": {
          "value": 0.8473715298287064,
          "lo": 0.8399738848348878,
          "hi": 0.8569233002961383
        }
      }
    },
    {
      "env": "folktables",
      "slice": "all",
      "system": "LFM2.5-VL-3B",
      "metrics": {
        "n": {
          "value": 15003.0,
          "lo": 14855.8,
          "hi": 15157.0
        },
        "ce": {
          "value": 0.6757370971930118,
          "lo": 0.6632898334485878,
          "hi": 0.6906677526353899
        },
        "kl": {
          "value": 0.6757370971930118,
          "lo": 0.6632898334485878,
          "hi": 0.6906677526353899
        },
        "jsd": {
          "value": 0.17894972031199147,
          "lo": 0.1762759338754224,
          "hi": 0.18216940431510975
        },
        "brier": {
          "value": 0.44920946997907557,
          "lo": 0.44049943112591494,
          "hi": 0.4595456620134782
        },
        "acc": {
          "value": 0.6783976538025728,
          "lo": 0.6702337979068954,
          "hi": 0.6859475904305837
        },
        "conf": {
          "value": 0.7932900791326173,
          "lo": 0.7909126609219911,
          "hi": 0.7954442834879843
        },
        "ece": {
          "value": 0.11489242533004451,
          "lo": 0.10689556856059901,
          "hi": 0.122851500909695
        },
        "reliability": {
          "value": 0.015857651914882486,
          "lo": 0.01431326500314493,
          "hi": 0.018158490729361654
        },
        "resolution": {
          "value": 0.009356267506864168,
          "lo": 0.008159686778144225,
          "hi": 0.010873125948180637
        },
        "uncertainty": {
          "value": 0.2181742771177374,
          "lo": 0.21542349156276827,
          "hi": 0.2210204522057745
        },
        "auroc": {
          "value": 0.6186877619818016,
          "lo": 0.609356710370421,
          "hi": 0.6281315833357176
        },
        "acc@80": {
          "value": 0.717380436593901,
          "lo": 0.7081394662921349,
          "hi": 0.7246021461791023
        }
      }
    },
    {
      "env": "folktables",
      "slice": "all",
      "system": "LFM2.5-VL-3B +T",
      "metrics": {
        "n": {
          "value": 15003.0,
          "lo": 14855.8,
          "hi": 15157.0
        },
        "ce": {
          "value": 0.609676541729926,
          "lo": 0.60345749975659,
          "hi": 0.6172599009517398
        },
        "kl": {
          "value": 0.609676541729926,
          "lo": 0.60345749975659,
          "hi": 0.6172599009517398
        },
        "jsd": {
          "value": 0.18367213312589367,
          "lo": 0.1820172407129862,
          "hi": 0.18564696915229778
        },
        "brier": {
          "value": 0.4202375772553369,
          "lo": 0.4146823417587002,
          "hi": 0.42692297904544924
        },
        "acc": {
          "value": 0.6783976538025728,
          "lo": 0.6702337979068954,
          "hi": 0.6859475904305837
        },
        "conf": {
          "value": 0.6819142642418174,
          "lo": 0.6801775258058251,
          "hi": 0.6835818231655569
        },
        "ece": {
          "value": 0.02830506018556748,
          "lo": 0.02470489534954933,
          "hi": 0.03731011481623176
        },
        "reliability": {
          "value": 0.0014208268752148375,
          "lo": 0.0010898369561207933,
          "hi": 0.002237282144687026
        },
        "resolution": {
          "value": 0.009356267506864168,
          "lo": 0.008159686778144225,
          "hi": 0.010873125948180637
        },
        "uncertainty": {
          "value": 0.2181742771177374,
          "lo": 0.21542349156276827,
          "hi": 0.2210204522057745
        },
        "auroc": {
          "value": 0.6186877619818016,
          "lo": 0.609356710370421,
          "hi": 0.6281315833357176
        },
        "acc@80": {
          "value": 0.717380436593901,
          "lo": 0.7081394662921349,
          "hi": 0.7246021461791023
        }
      }
    },
    {
      "env": "mcq",
      "slice": "all",
      "system": "LFM2.5-VL-3B",
      "metrics": {
        "n": {
          "value": 5064.0,
          "lo": 5064.0,
          "hi": 5064.0
        },
        "ce": {
          "value": 0.5759872670241797,
          "lo": 0.5329584569816519,
          "hi": 0.6265682381052227
        },
        "kl": {
          "value": 0.5759872670241797,
          "lo": 0.5329584569816519,
          "hi": 0.6265682381052227
        },
        "jsd": {
          "value": 0.08353110019922566,
          "lo": 0.07820357437215822,
          "hi": 0.08958165164902972
        },
        "brier": {
          "value": 0.22348060549884158,
          "lo": 0.20822744166423812,
          "hi": 0.2404070146904414
        },
        "acc": {
          "value": 0.868088467614534,
          "lo": 0.858901066350711,
          "hi": 0.8766834518167457
        },
        "conf": {
          "value": 0.9615081339474255,
          "lo": 0.9586090640788736,
          "hi": 0.9641287516267002
        },
        "ece": {
          "value": 0.09343844024232026,
          "lo": 0.08509777928960008,
          "hi": 0.1014441097433286
        },
        "reliability": {
          "value": 0.019725029048978552,
          "lo": 0.016627496437260518,
          "hi": 0.023582789751029707
        },
        "resolution": {
          "value": 0.028799536189915965,
          "lo": 0.02509962954635008,
          "hi": 0.03310776419602363
        },
        "uncertainty": {
          "value": 0.11451088000918419,
          "lo": 0.1081095674029609,
          "hi": 0.12119001484786204
        },
        "auroc": {
          "value": 0.8771137206932813,
          "lo": 0.866687062555456,
          "hi": 0.8878373819364807
        },
        "acc@80": {
          "value": 0.9427301900765244,
          "lo": 0.9355714638360898,
          "hi": 0.9497716613181931
        }
      }
    },
    {
      "env": "mcq",
      "slice": "all",
      "system": "LFM2.5-VL-3B +T",
      "metrics": {
        "n": {
          "value": 5064.0,
          "lo": 5064.0,
          "hi": 5064.0
        },
        "ce": {
          "value": 0.3906034227460842,
          "lo": 0.37137167691447304,
          "hi": 0.41206851196394295
        },
        "kl": {
          "value": 0.3906034227460842,
          "lo": 0.37137167691447304,
          "hi": 0.41206851196394295
        },
        "jsd": {
          "value": 0.10002890622380546,
          "lo": 0.09610960090860976,
          "hi": 0.10440735944956796
        },
        "brier": {
          "value": 0.1937977066232963,
          "lo": 0.18263927510816466,
          "hi": 0.20566609163289462
        },
        "acc": {
          "value": 0.868088467614534,
          "lo": 0.858901066350711,
          "hi": 0.8766834518167457
        },
        "conf": {
          "value": 0.8558151643390487,
          "lo": 0.8511140746036796,
          "hi": 0.8593811069517472
        },
        "ece": {
          "value": 0.0305649517438755,
          "lo": 0.026309415051930726,
          "hi": 0.039549654257689214
        },
        "reliability": {
          "value": 0.0010577246490419798,
          "lo": 0.0008757950986678834,
          "hi": 0.0019207445946941867
        },
        "resolution": {
          "value": 0.029804711091153086,
          "lo": 0.026167586198610596,
          "hi": 0.03420775328193042
        },
        "uncertainty": {
          "value": 0.11451088000918419,
          "lo": 0.1081095674029609,
          "hi": 0.12119001484786204
        },
        "auroc": {
          "value": 0.8805730781385364,
          "lo": 0.8702694072406397,
          "hi": 0.8911781380761858
        },
        "acc@80": {
          "value": 0.9449518637373489,
          "lo": 0.9370525796099728,
          "hi": 0.9511231794618613
        }
      }
    },
    {
      "env": "stsb",
      "slice": "all",
      "system": "LFM2.5-VL-3B",
      "metrics": {
        "n": {
          "value": 3676.0,
          "lo": 3150.95,
          "hi": 4793.674999999996
        },
        "ce": {
          "value": 2.6978885169981255,
          "lo": 2.5657299873689126,
          "hi": 2.8476372111272323
        },
        "kl": {
          "value": 2.2955846941358695,
          "lo": 2.15679378594023,
          "hi": 2.4465032494598424
        },
        "jsd": {
          "value": 0.3275760175289124,
          "lo": 0.3162216290936204,
          "hi": 0.3390716580405639
        },
        "brier": {
          "value": 0.6919414239196017,
          "lo": 0.6638624637074109,
          "hi": 0.7198471346121803
        },
        "acc": {
          "value": 0.3520130576713819,
          "lo": 0.33738188605734065,
          "hi": 0.36734544487166837
        },
        "conf": {
          "value": 0.7248023496748184,
          "lo": 0.7112682362621677,
          "hi": 0.7419135614849961
        },
        "ece": {
          "value": 0.3727892920034365,
          "lo": 0.34990632614727657,
          "hi": 0.39600605273239675
        },
        "reliability": {
          "value": 0.1495211764503258,
          "lo": 0.13412612551018585,
          "hi": 0.1686786364778362
        },
        "resolution": {
          "value": 0.007239650807423352,
          "lo": 0.005789911237959588,
          "hi": 0.010394805507148573
        },
        "uncertainty": {
          "value": 0.22809986490022627,
          "lo": 0.22355534383864203,
          "hi": 0.23240276779454366
        },
        "auroc": {
          "value": 0.5737491191665466,
          "lo": 0.5525762701641901,
          "hi": 0.5929816932597282
        },
        "acc@80": {
          "value": 0.3651819109146549,
          "lo": 0.34728186363682,
          "hi": 0.3814624881124586
        }
      }
    },
    {
      "env": "stsb",
      "slice": "all",
      "system": "LFM2.5-VL-3B +T",
      "metrics": {
        "n": {
          "value": 3676.0,
          "lo": 3150.95,
          "hi": 4793.674999999996
        },
        "ce": {
          "value": 1.5758103744712262,
          "lo": 1.552264621029743,
          "hi": 1.6003796429076587
        },
        "kl": {
          "value": 1.1735065516089704,
          "lo": 1.1455278721099897,
          "hi": 1.2018428465852689
        },
        "jsd": {
          "value": 0.3105009562586476,
          "lo": 0.3048929847451175,
          "hi": 0.31570691507361504
        },
        "brier": {
          "value": 0.48641483403122787,
          "lo": 0.4735061300168612,
          "hi": 0.49741827930201415
        },
        "acc": {
          "value": 0.3520130576713819,
          "lo": 0.33738188605734065,
          "hi": 0.36734544487166837
        },
        "conf": {
          "value": 0.3528447149650681,
          "lo": 0.3482181155596457,
          "hi": 0.35906599390698796
        },
        "ece": {
          "value": 0.04800463541314871,
          "lo": 0.041971453587293386,
          "hi": 0.07054283841584154
        },
        "reliability": {
          "value": 0.0058868450744030375,
          "lo": 0.0038970035004699415,
          "hi": 0.008971929159702999
        },
        "resolution": {
          "value": 0.012971840177239595,
          "lo": 0.009013087833499103,
          "hi": 0.017507180870015102
        },
        "uncertainty": {
          "value": 0.22809986490022627,
          "lo": 0.22355534383864203,
          "hi": 0.23240276779454366
        },
        "auroc": {
          "value": 0.6127025592510548,
          "lo": 0.5903542080481157,
          "hi": 0.633554195636134
        },
        "acc@80": {
          "value": 0.37810268616116965,
          "lo": 0.3606827313509045,
          "hi": 0.39444786464976245
        }
      }
    },
    {
      "env": "yelp",
      "slice": "all",
      "system": "LFM2.5-VL-3B",
      "metrics": {
        "n": {
          "value": 5000.0,
          "lo": 5000.0,
          "hi": 5000.0
        },
        "ce": {
          "value": 1.3860599847549677,
          "lo": 1.3290156657155492,
          "hi": 1.4319837141033827
        },
        "kl": {
          "value": 1.3860599847549677,
          "lo": 1.3290156657155492,
          "hi": 1.4319837141033827
        },
        "jsd": {
          "value": 0.24637008755393536,
          "lo": 0.23853545745746285,
          "hi": 0.2536433407530511
        },
        "brier": {
          "value": 0.6418344653090541,
          "lo": 0.6195384048344682,
          "hi": 0.6631553474924564
        },
        "acc": {
          "value": 0.5928,
          "lo": 0.578695,
          "hi": 0.6074
        },
        "conf": {
          "value": 0.8411075247108937,
          "lo": 0.8362002471122145,
          "hi": 0.8460747388541698
        },
        "ece": {
          "value": 0.24830752471089362,
          "lo": 0.2347119203391671,
          "hi": 0.26253560655146835
        },
        "reliability": {
          "value": 0.07551728217756373,
          "lo": 0.06811680137418837,
          "hi": 0.08436615957078013
        },
        "resolution": {
          "value": 0.021735186667385942,
          "lo": 0.018947388334327648,
          "hi": 0.025708930635887053
        },
        "uncertainty": {
          "value": 0.24138816,
          "lo": 0.23846524,
          "hi": 0.24380708699999998
        },
        "auroc": {
          "value": 0.6629740249066068,
          "lo": 0.6476209612008791,
          "hi": 0.6784682418750948
        },
        "acc@80": {
          "value": 0.626,
          "lo": 0.61036875,
          "hi": 0.64305
        }
      }
    },
    {
      "env": "yelp",
      "slice": "all",
      "system": "LFM2.5-VL-3B +T",
      "metrics": {
        "n": {
          "value": 5000.0,
          "lo": 5000.0,
          "hi": 5000.0
        },
        "ce": {
          "value": 0.9724022468050197,
          "lo": 0.9483220949408991,
          "hi": 0.9937898554692086
        },
        "kl": {
          "value": 0.9724022468050197,
          "lo": 0.9483220949408991,
          "hi": 0.9937898554692086
        },
        "jsd": {
          "value": 0.2549887608925926,
          "lo": 0.2497078453139047,
          "hi": 0.2595903865139499
        },
        "brier": {
          "value": 0.5386607794010954,
          "lo": 0.5238223524042669,
          "hi": 0.5520623252485813
        },
        "acc": {
          "value": 0.5928,
          "lo": 0.578695,
          "hi": 0.6074
        },
        "conf": {
          "value": 0.6407830723416755,
          "lo": 0.6355987860389466,
          "hi": 0.6457528767574988
        },
        "ece": {
          "value": 0.06264510225858502,
          "lo": 0.054522714912741445,
          "hi": 0.07988027057516923
        },
        "reliability": {
          "value": 0.005896093809502374,
          "lo": 0.004641341937083263,
          "hi": 0.008890860101156363
        },
        "resolution": {
          "value": 0.020065581733110356,
          "lo": 0.017560380981010818,
          "hi": 0.02410640542155473
        },
        "uncertainty": {
          "value": 0.24138816,
          "lo": 0.23846524,
          "hi": 0.24380708699999998
        },
        "auroc": {
          "value": 0.6625109533127059,
          "lo": 0.6478243237246345,
          "hi": 0.6778121607764245
        },
        "acc@80": {
          "value": 0.63075,
          "lo": 0.61384375,
          "hi": 0.64675
        }
      }
    },
    {
      "env": "boolq",
      "slice": "all",
      "system": "zs-LFM2.5-VL-3B-prefill",
      "metrics": {
        "n": {
          "value": 8129.0,
          "lo": 7984.475,
          "hi": 8270.575
        },
        "ce": {
          "value": 0.5816780320630307,
          "lo": 0.5487705254824645,
          "hi": 0.6153245192490632
        },
        "kl": {
          "value": 0.5816780320630307,
          "lo": 0.5487705254824645,
          "hi": 0.6153245192490632
        },
        "jsd": {
          "value": 0.10986828522452227,
          "lo": 0.1050696462488618,
          "hi": 0.1150266473853312
        },
        "brier": {
          "value": 0.2965075321276258,
          "lo": 0.2821260469188488,
          "hi": 0.3115855130827731
        },
        "acc": {
          "value": 0.8144913273465371,
          "lo": 0.8049283683473497,
          "hi": 0.8234987300629023
        },
        "conf": {
          "value": 0.9305177249875012,
          "lo": 0.9276152868000491,
          "hi": 0.9331465415011992
        },
        "ece": {
          "value": 0.11602639764096409,
          "lo": 0.10767501255807296,
          "hi": 0.12471576912306866
        },
        "reliability": {
          "value": 0.016211774414308433,
          "lo": 0.013994393412881936,
          "hi": 0.019032058298974353
        },
        "resolution": {
          "value": 0.018600505289024332,
          "lo": 0.016417993663277192,
          "hi": 0.021259464248592964
        },
        "uncertainty": {
          "value": 0.15109520502381327,
          "lo": 0.14534857005625082,
          "hi": 0.15701869003971955
        },
        "auroc": {
          "value": 0.7468794030888777,
          "lo": 0.7331697018418196,
          "hi": 0.7595890184078872
        },
        "acc@80": {
          "value": 0.8699061971397817,
          "lo": 0.8604725626403062,
          "hi": 0.8787990101196874
        }
      }
    },
    {
      "env": "boolq",
      "slice": "all",
      "system": "zs-LFM2.5-VL-3B-prefill +T",
      "metrics": {
        "n": {
          "value": 8129.0,
          "lo": 7984.475,
          "hi": 8270.575
        },
        "ce": {
          "value": 0.42064213095297337,
          "lo": 0.40627805325118327,
          "hi": 0.43623354700168643
        },
        "kl": {
          "value": 0.42064213095297337,
          "lo": 0.40627805325118327,
          "hi": 0.43623354700168643
        },
        "jsd": {
          "value": 0.11765313022620086,
          "lo": 0.1142472051321885,
          "hi": 0.12136347298891204
        },
        "brier": {
          "value": 0.2654743249822674,
          "lo": 0.25476302978534976,
          "hi": 0.2768470924704069
        },
        "acc": {
          "value": 0.8144913273465371,
          "lo": 0.8049283683473497,
          "hi": 0.8234987300629023
        },
        "conf": {
          "value": 0.8300160892799145,
          "lo": 0.8266725368018208,
          "hi": 0.832904834456254
        },
        "ece": {
          "value": 0.017711542259073728,
          "lo": 0.01568471163053336,
          "hi": 0.0289449105887779
        },
        "reliability": {
          "value": 0.0006015042336070265,
          "lo": 0.00040831449968761535,
          "hi": 0.001444573067852351
        },
        "resolution": {
          "value": 0.018600505289024332,
          "lo": 0.016417993663277192,
          "hi": 0.021259464248592964
        },
        "uncertainty": {
          "value": 0.15109520502381327,
          "lo": 0.14534857005625082,
          "hi": 0.15701869003971955
        },
        "auroc": {
          "value": 0.7468794030888777,
          "lo": 0.7331697018418196,
          "hi": 0.7595890184078872
        },
        "acc@80": {
          "value": 0.8699061971397817,
          "lo": 0.8604725626403062,
          "hi": 0.8787990101196874
        }
      }
    },
    {
      "env": "chaosnli",
      "slice": "all",
      "system": "zs-LFM2.5-VL-3B-prefill",
      "metrics": {
        "n": {
          "value": 2763.0,
          "lo": 2732.475,
          "hi": 2798.525
        },
        "ce": {
          "value": 1.4204363394402675,
          "lo": 1.3710480649425658,
          "hi": 1.46618904115787
        },
        "kl": {
          "value": 0.8900762693710054,
          "lo": 0.8454883049121936,
          "hi": 0.9338853046818825
        },
        "jsd": {
          "value": 0.12223944325440186,
          "lo": 0.11642119374214775,
          "hi": 0.127179793442689
        },
        "brier": {
          "value": 0.3237633895884616,
          "lo": 0.3063289695870303,
          "hi": 0.3388995012898196
        },
        "acc": {
          "value": 0.664133188563156,
          "lo": 0.6471050323227849,
          "hi": 0.6809085836106766
        },
        "conf": {
          "value": 0.8941022409095827,
          "lo": 0.8889790748660467,
          "hi": 0.899864060352879
        },
        "ece": {
          "value": 0.22996905234642664,
          "lo": 0.214183808314927,
          "hi": 0.24590967789885282
        },
        "reliability": {
          "value": 0.058616589135346245,
          "lo": 0.05185238690359973,
          "hi": 0.06779369685426515
        },
        "resolution": {
          "value": 0.02073116596346457,
          "lo": 0.01743926586421297,
          "hi": 0.026394245867025382
        },
        "uncertainty": {
          "value": 0.2230602964120915,
          "lo": 0.21727207848417857,
          "hi": 0.22836008361661222
        },
        "auroc": {
          "value": 0.679469193366532,
          "lo": 0.6591451012025751,
          "hi": 0.6992481806264657
        },
        "acc@80": {
          "value": 0.7167420814479638,
          "lo": 0.6975548414267926,
          "hi": 0.7350212868900042
        }
      }
    },
    {
      "env": "chaosnli",
      "slice": "all",
      "system": "zs-LFM2.5-VL-3B-prefill +T",
      "metrics": {
        "n": {
          "value": 2763.0,
          "lo": 2732.475,
          "hi": 2798.525
        },
        "ce": {
          "value": 0.7949680342778785,
          "lo": 0.7820818091143783,
          "hi": 0.8070480612599928
        },
        "kl": {
          "value": 0.2646079642086167,
          "lo": 0.25460557149217655,
          "hi": 0.2749783513874871
        },
        "jsd": {
          "value": 0.07147578627026423,
          "lo": 0.06895103092046842,
          "hi": 0.07409254429129038
        },
        "brier": {
          "value": 0.17245520194055936,
          "lo": 0.16471963076202234,
          "hi": 0.18086821914340606
        },
        "acc": {
          "value": 0.664133188563156,
          "lo": 0.6471050323227849,
          "hi": 0.6809085836106766
        },
        "conf": {
          "value": 0.6241185789225275,
          "lo": 0.6198024413713349,
          "hi": 0.6292444607739884
        },
        "ece": {
          "value": 0.056334159207648,
          "lo": 0.045386576939123006,
          "hi": 0.0740956807468349
        },
        "reliability": {
          "value": 0.004457731467115396,
          "lo": 0.003090377892814316,
          "hi": 0.0074727726597290005
        },
        "resolution": {
          "value": 0.018693482513854446,
          "lo": 0.015458552568498226,
          "hi": 0.02357529165847566
        },
        "uncertainty": {
          "value": 0.2230602964120915,
          "lo": 0.21727207848417857,
          "hi": 0.22836008361661222
        },
        "auroc": {
          "value": 0.6649402776472799,
          "lo": 0.6420207504509805,
          "hi": 0.686683594932958
        },
        "acc@80": {
          "value": 0.7149321266968326,
          "lo": 0.6973889241402683,
          "hi": 0.7339677960187573
        }
      }
    },
    {
      "env": "civilcomments",
      "slice": "all",
      "system": "zs-LFM2.5-VL-3B-prefill",
      "metrics": {
        "n": {
          "value": 10581.0,
          "lo": 10261.425,
          "hi": 10904.525
        },
        "ce": {
          "value": 0.43727784798305874,
          "lo": 0.42678145217047764,
          "hi": 0.4476415325693049
        },
        "kl": {
          "value": 0.20579503921371767,
          "lo": 0.1987235586387011,
          "hi": 0.21235390816668306
        },
        "jsd": {
          "value": 0.05116968131305278,
          "lo": 0.049437446831806064,
          "hi": 0.052595588753110684
        },
        "brier": {
          "value": 0.11855361891618092,
          "lo": 0.11343503373455538,
          "hi": 0.12313664944780268
        },
        "acc": {
          "value": 0.8596540969662603,
          "lo": 0.8527050254805175,
          "hi": 0.8673564438680864
        },
        "conf": {
          "value": 0.8579861902861635,
          "lo": 0.8548736530748258,
          "hi": 0.8611949410419385
        },
        "ece": {
          "value": 0.0115804810761031,
          "lo": 0.009709700819945096,
          "hi": 0.020409191830464038
        },
        "reliability": {
          "value": 0.00031894320172582623,
          "lo": 0.0001891966330327208,
          "hi": 0.000863846155247233
        },
        "resolution": {
          "value": 0.01624237040360173,
          "lo": 0.014398505478895042,
          "hi": 0.018297819452531085
        },
        "uncertainty": {
          "value": 0.12064893053538385,
          "lo": 0.11504924303043536,
          "hi": 0.1255991644744319
        },
        "auroc": {
          "value": 0.79042595405832,
          "lo": 0.7795293025825059,
          "hi": 0.8014689978834265
        },
        "acc@80": {
          "value": 0.9135262847017129,
          "lo": 0.9061254603062078,
          "hi": 0.920151691035287
        }
      }
    },
    {
      "env": "civilcomments",
      "slice": "all",
      "system": "zs-LFM2.5-VL-3B-prefill +T",
      "metrics": {
        "n": {
          "value": 10581.0,
          "lo": 10261.425,
          "hi": 10904.525
        },
        "ce": {
          "value": 0.422683145843368,
          "lo": 0.4138486310025811,
          "hi": 0.4310942217711616
        },
        "kl": {
          "value": 0.19120033707402692,
          "lo": 0.18542010246649468,
          "hi": 0.19586281686802368
        },
        "jsd": {
          "value": 0.053667711340155445,
          "lo": 0.05219516817925039,
          "hi": 0.05481715442924367
        },
        "brier": {
          "value": 0.11366267360512225,
          "lo": 0.10908920692600414,
          "hi": 0.1175331613875349
        },
        "acc": {
          "value": 0.8596540969662603,
          "lo": 0.8527050254805175,
          "hi": 0.8673564438680864
        },
        "conf": {
          "value": 0.8151815127133928,
          "lo": 0.8119748811039861,
          "hi": 0.8183424007286001
        },
        "ece": {
          "value": 0.04447258425286756,
          "lo": 0.038576093724351536,
          "hi": 0.05067963280226556
        },
        "reliability": {
          "value": 0.0021773724847827768,
          "lo": 0.0017861691523396307,
          "hi": 0.003071245325309536
        },
        "resolution": {
          "value": 0.01624237040360173,
          "lo": 0.014398505478895042,
          "hi": 0.018297819452531085
        },
        "uncertainty": {
          "value": 0.12064893053538385,
          "lo": 0.11504924303043536,
          "hi": 0.1255991644744319
        },
        "auroc": {
          "value": 0.79042595405832,
          "lo": 0.7795293025825059,
          "hi": 0.8014689978834265
        },
        "acc@80": {
          "value": 0.9135262847017129,
          "lo": 0.9061254603062078,
          "hi": 0.920151691035287
        }
      }
    },
    {
      "env": "folktables",
      "slice": "all",
      "system": "zs-LFM2.5-VL-3B-prefill",
      "metrics": {
        "n": {
          "value": 15003.0,
          "lo": 14855.8,
          "hi": 15157.0
        },
        "ce": {
          "value": 0.9620101471971333,
          "lo": 0.9402995881311484,
          "hi": 0.9863653242467205
        },
        "kl": {
          "value": 0.9620101471971333,
          "lo": 0.9402995881311484,
          "hi": 0.9863653242467205
        },
        "jsd": {
          "value": 0.1987697950575701,
          "lo": 0.19457230876065962,
          "hi": 0.20330264801141593
        },
        "brier": {
          "value": 0.5668222757523853,
          "lo": 0.5533690394685322,
          "hi": 0.5807961863890092
        },
        "acc": {
          "value": 0.6554689062187562,
          "lo": 0.6476506602729011,
          "hi": 0.6634009913565764
        },
        "conf": {
          "value": 0.9244254015488902,
          "lo": 0.9230365547995866,
          "hi": 0.9257004334141172
        },
        "ece": {
          "value": 0.26895649533013394,
          "lo": 0.2608665257868461,
          "hi": 0.2770311776931512
        },
        "reliability": {
          "value": 0.0779897735496795,
          "lo": 0.07362743547269779,
          "hi": 0.08285164159173998
        },
        "resolution": {
          "value": 0.020396192527323317,
          "lo": 0.01845828721270402,
          "hi": 0.02260715738277575
        },
        "uncertainty": {
          "value": 0.2258294191991436,
          "lo": 0.22330011591143803,
          "hi": 0.22819928249323598
        },
        "auroc": {
          "value": 0.6630820901485849,
          "lo": 0.6535759412768337,
          "hi": 0.6720372414084176
        },
        "acc@80": {
          "value": 0.7175470754874188,
          "lo": 0.7077399384103128,
          "hi": 0.7245951236863847
        }
      }
    },
    {
      "env": "folktables",
      "slice": "all",
      "system": "zs-LFM2.5-VL-3B-prefill +T",
      "metrics": {
        "n": {
          "value": 15003.0,
          "lo": 14855.8,
          "hi": 15157.0
        },
        "ce": {
          "value": 0.617889507782528,
          "lo": 0.6118568964850071,
          "hi": 0.6249296516130413
        },
        "kl": {
          "value": 0.617889507782528,
          "lo": 0.6118568964850071,
          "hi": 0.6249296516130413
        },
        "jsd": {
          "value": 0.1869655766677398,
          "lo": 0.18535809436615105,
          "hi": 0.18885473665697913
        },
        "brier": {
          "value": 0.4285013151213943,
          "lo": 0.42284947531250267,
          "hi": 0.43501332745244986
        },
        "acc": {
          "value": 0.6554689062187562,
          "lo": 0.6476506602729011,
          "hi": 0.6634009913565764
        },
        "conf": {
          "value": 0.6880828157315765,
          "lo": 0.6869715589806281,
          "hi": 0.6891711087324444
        },
        "ece": {
          "value": 0.06847379056915753,
          "lo": 0.0614792476384789,
          "hi": 0.07600070537683982
        },
        "reliability": {
          "value": 0.009036866185085667,
          "lo": 0.00782397103727989,
          "hi": 0.010700940911057373
        },
        "resolution": {
          "value": 0.020396192527323317,
          "lo": 0.01845828721270402,
          "hi": 0.02260715738277575
        },
        "uncertainty": {
          "value": 0.2258294191991436,
          "lo": 0.22330011591143803,
          "hi": 0.22819928249323598
        },
        "auroc": {
          "value": 0.6630820901485849,
          "lo": 0.6535759412768337,
          "hi": 0.6720372414084176
        },
        "acc@80": {
          "value": 0.7175470754874188,
          "lo": 0.7077399384103128,
          "hi": 0.7245951236863847
        }
      }
    },
    {
      "env": "mcq",
      "slice": "all",
      "system": "zs-LFM2.5-VL-3B-prefill",
      "metrics": {
        "n": {
          "value": 5064.0,
          "lo": 5064.0,
          "hi": 5064.0
        },
        "ce": {
          "value": 0.53918624449217,
          "lo": 0.49918554138227583,
          "hi": 0.5816754544353302
        },
        "kl": {
          "value": 0.53918624449217,
          "lo": 0.49918554138227583,
          "hi": 0.5816754544353302
        },
        "jsd": {
          "value": 0.0828151557503379,
          "lo": 0.07740562729541314,
          "hi": 0.08866953235726405
        },
        "brier": {
          "value": 0.2200081923930159,
          "lo": 0.20467040691603539,
          "hi": 0.23641613618774399
        },
        "acc": {
          "value": 0.868088467614534,
          "lo": 0.8587035939968405,
          "hi": 0.8769747235387045
        },
        "conf": {
          "value": 0.9587305887038768,
          "lo": 0.955746518491398,
          "hi": 0.9612954949833582
        },
        "ece": {
          "value": 0.09069410544977557,
          "lo": 0.08256884661310399,
          "hi": 0.09929659935744335
        },
        "reliability": {
          "value": 0.01824473815330227,
          "lo": 0.015307869170038992,
          "hi": 0.021990593540975493
        },
        "resolution": {
          "value": 0.029541577003898572,
          "lo": 0.025870115631190058,
          "hi": 0.03445812620111184
        },
        "uncertainty": {
          "value": 0.11451088000918419,
          "lo": 0.10789005781291729,
          "hi": 0.1213317219292893
        },
        "auroc": {
          "value": 0.8802924746503353,
          "lo": 0.869610276682273,
          "hi": 0.8906437648543074
        },
        "acc@80": {
          "value": 0.9439644532214269,
          "lo": 0.9363120217230313,
          "hi": 0.9516168847198223
        }
      }
    },
    {
      "env": "mcq",
      "slice": "all",
      "system": "zs-LFM2.5-VL-3B-prefill +T",
      "metrics": {
        "n": {
          "value": 5064.0,
          "lo": 5064.0,
          "hi": 5064.0
        },
        "ce": {
          "value": 0.38602727384245317,
          "lo": 0.3674236940481484,
          "hi": 0.4071443140733702
        },
        "kl": {
          "value": 0.38602727384245317,
          "lo": 0.3674236940481484,
          "hi": 0.4071443140733702
        },
        "jsd": {
          "value": 0.09900290314036149,
          "lo": 0.09492639363834217,
          "hi": 0.10347017607384866
        },
        "brier": {
          "value": 0.1923688280406221,
          "lo": 0.1813023317296054,
          "hi": 0.20459157811467585
        },
        "acc": {
          "value": 0.868088467614534,
          "lo": 0.8587035939968405,
          "hi": 0.8769747235387045
        },
        "conf": {
          "value": 0.8574229183584634,
          "lo": 0.8528227884766457,
          "hi": 0.8610242711011652
        },
        "ece": {
          "value": 0.03227782449687891,
          "lo": 0.02757931630534842,
          "hi": 0.04139280942821726
        },
        "reliability": {
          "value": 0.0011589286567429886,
          "lo": 0.0008960357913958104,
          "hi": 0.002189673488796682
        },
        "resolution": {
          "value": 0.029976965872117813,
          "lo": 0.026465285875255894,
          "hi": 0.03492715316660374
        },
        "uncertainty": {
          "value": 0.11451088000918419,
          "lo": 0.10789005781291729,
          "hi": 0.1213317219292893
        },
        "auroc": {
          "value": 0.8834581178861567,
          "lo": 0.8733877344785387,
          "hi": 0.8935237529435539
        },
        "acc@80": {
          "value": 0.9454455689953098,
          "lo": 0.9380399901258949,
          "hi": 0.9521105899777833
        }
      }
    },
    {
      "env": "stsb",
      "slice": "all",
      "system": "zs-LFM2.5-VL-3B-prefill",
      "metrics": {
        "n": {
          "value": 3676.0,
          "lo": 3150.95,
          "hi": 4793.674999999996
        },
        "ce": {
          "value": 2.7571148750759926,
          "lo": 2.634763394031497,
          "hi": 2.8926179822485656
        },
        "kl": {
          "value": 2.354811052213737,
          "lo": 2.2283260410334957,
          "hi": 2.4908382476713715
        },
        "jsd": {
          "value": 0.33952460202990814,
          "lo": 0.3276339706798127,
          "hi": 0.35131355014479676
        },
        "brier": {
          "value": 0.7369648478638614,
          "lo": 0.707818783439776,
          "hi": 0.7640002657837547
        },
        "acc": {
          "value": 0.3291621327529924,
          "lo": 0.3153417244862836,
          "hi": 0.3446662586052942
        },
        "conf": {
          "value": 0.7506276568776894,
          "lo": 0.7392975738806389,
          "hi": 0.7643049989491109
        },
        "ece": {
          "value": 0.42146552412469707,
          "lo": 0.40124266760385313,
          "hi": 0.44190863029437977
        },
        "reliability": {
          "value": 0.1858167116582256,
          "lo": 0.17023990225473562,
          "hi": 0.2029554849167609
        },
        "resolution": {
          "value": 0.005736625759717607,
          "lo": 0.004384957131903068,
          "hi": 0.009514040670213124
        },
        "uncertainty": {
          "value": 0.2208144231144938,
          "lo": 0.21590132128117429,
          "hi": 0.2258713780811604
        },
        "auroc": {
          "value": 0.5797091686607281,
          "lo": 0.5610394723800967,
          "hi": 0.5973409348331422
        },
        "acc@80": {
          "value": 0.3498809928595716,
          "lo": 0.33342935112189204,
          "hi": 0.3666129559917872
        }
      }
    },
    {
      "env": "stsb",
      "slice": "all",
      "system": "zs-LFM2.5-VL-3B-prefill +T",
      "metrics": {
        "n": {
          "value": 3676.0,
          "lo": 3150.95,
          "hi": 4793.674999999996
        },
        "ce": {
          "value": 1.5891963814976409,
          "lo": 1.5669328002567837,
          "hi": 1.6120440647134835
        },
        "kl": {
          "value": 1.1868925586353853,
          "lo": 1.1599661816058862,
          "hi": 1.2138312690286734
        },
        "jsd": {
          "value": 0.3146720961655134,
          "lo": 0.3091232327268351,
          "hi": 0.3198014719773035
        },
        "brier": {
          "value": 0.49317121724101637,
          "lo": 0.4802501732716859,
          "hi": 0.5042819420802195
        },
        "acc": {
          "value": 0.3291621327529924,
          "lo": 0.3153417244862836,
          "hi": 0.3446662586052942
        },
        "conf": {
          "value": 0.3511478778518168,
          "lo": 0.3477794827058472,
          "hi": 0.3554372821668198
        },
        "ece": {
          "value": 0.05635346653229976,
          "lo": 0.04942946070230882,
          "hi": 0.07853436982282196
        },
        "reliability": {
          "value": 0.006045812371941183,
          "lo": 0.004487231318838066,
          "hi": 0.009358210914887517
        },
        "resolution": {
          "value": 0.011532073268055585,
          "lo": 0.008924448325187958,
          "hi": 0.016059102874527314
        },
        "uncertainty": {
          "value": 0.2208144231144938,
          "lo": 0.21590132128117429,
          "hi": 0.2258713780811604
        },
        "auroc": {
          "value": 0.6206665192066652,
          "lo": 0.5980069035086956,
          "hi": 0.6401835762902832
        },
        "acc@80": {
          "value": 0.3594015640938456,
          "lo": 0.34184327632218964,
          "hi": 0.3758668855010105
        }
      }
    },
    {
      "env": "yelp",
      "slice": "all",
      "system": "zs-LFM2.5-VL-3B-prefill",
      "metrics": {
        "n": {
          "value": 5000.0,
          "lo": 5000.0,
          "hi": 5000.0
        },
        "ce": {
          "value": 1.2918572553652283,
          "lo": 1.2399163291387825,
          "hi": 1.333798228136019
        },
        "kl": {
          "value": 1.2918572553652283,
          "lo": 1.2399163291387825,
          "hi": 1.333798228136019
        },
        "jsd": {
          "value": 0.24252614524047386,
          "lo": 0.23506189151301068,
          "hi": 0.24962484157285553
        },
        "brier": {
          "value": 0.627556481522769,
          "lo": 0.6061041358647522,
          "hi": 0.6478518299527196
        },
        "acc": {
          "value": 0.5952,
          "lo": 0.580695,
          "hi": 0.6106
        },
        "conf": {
          "value": 0.8340039846897125,
          "lo": 0.8292275972224772,
          "hi": 0.8386646578364074
        },
        "ece": {
          "value": 0.23880398468971253,
          "lo": 0.22476129868015646,
          "hi": 0.25315469246715305
        },
        "reliability": {
          "value": 0.07083242179295059,
          "lo": 0.06359410323551529,
          "hi": 0.07945601400217328
        },
        "resolution": {
          "value": 0.02347230412776531,
          "lo": 0.02062576803127584,
          "hi": 0.027707121829315607
        },
        "uncertainty": {
          "value": 0.24093696,
          "lo": 0.23776764,
          "hi": 0.243488307
        },
        "auroc": {
          "value": 0.6694239024182923,
          "lo": 0.6536613432275953,
          "hi": 0.683653191239962
        },
        "acc@80": {
          "value": 0.63175,
          "lo": 0.6155,
          "hi": 0.6495125
        }
      }
    },
    {
      "env": "yelp",
      "slice": "all",
      "system": "zs-LFM2.5-VL-3B-prefill +T",
      "metrics": {
        "n": {
          "value": 5000.0,
          "lo": 5000.0,
          "hi": 5000.0
        },
        "ce": {
          "value": 0.9605660163022847,
          "lo": 0.9369846035904397,
          "hi": 0.9816685557479368
        },
        "kl": {
          "value": 0.9605660163022847,
          "lo": 0.9369846035904397,
          "hi": 0.9816685557479368
        },
        "jsd": {
          "value": 0.2516838502783082,
          "lo": 0.24637970558942096,
          "hi": 0.25627033242451014
        },
        "brier": {
          "value": 0.5332790437612978,
          "lo": 0.5184867031910769,
          "hi": 0.5463571151359354
        },
        "acc": {
          "value": 0.5952,
          "lo": 0.580695,
          "hi": 0.6106
        },
        "conf": {
          "value": 0.6486309407859533,
          "lo": 0.6437934317585855,
          "hi": 0.6531680733905414
        },
        "ece": {
          "value": 0.06880398054217841,
          "lo": 0.058410055797531384,
          "hi": 0.08362566469544623
        },
        "reliability": {
          "value": 0.007127271545934169,
          "lo": 0.005628349175403814,
          "hi": 0.010483590238076811
        },
        "resolution": {
          "value": 0.022972801829853808,
          "lo": 0.019711298467703607,
          "hi": 0.02692298937206552
        },
        "uncertainty": {
          "value": 0.24093696,
          "lo": 0.23776764,
          "hi": 0.243488307
        },
        "auroc": {
          "value": 0.6694089607505631,
          "lo": 0.6543323056358303,
          "hi": 0.6836065841514896
        },
        "acc@80": {
          "value": 0.63325,
          "lo": 0.61686875,
          "hi": 0.6505375
        }
      }
    },
    {
      "env": "boolq",
      "slice": "all",
      "system": "mix7-1.6B-s0",
      "metrics": {
        "n": {
          "value": 8129.0,
          "lo": 7984.475,
          "hi": 8270.575
        },
        "ce": {
          "value": 0.5082761310374166,
          "lo": 0.49592209899738315,
          "hi": 0.5199124539145769
        },
        "kl": {
          "value": 0.5082761310374166,
          "lo": 0.49592209899738315,
          "hi": 0.5199124539145769
        },
        "jsd": {
          "value": 0.1506662270607344,
          "lo": 0.1475305328937042,
          "hi": 0.1537566314506511
        },
        "brier": {
          "value": 0.34082382639915726,
          "lo": 0.3305859817855617,
          "hi": 0.35073283205382355
        },
        "acc": {
          "value": 0.7293640054127198,
          "lo": 0.7184596840938764,
          "hi": 0.7395318200163514
        },
        "conf": {
          "value": 0.7481786015949092,
          "lo": 0.7447704591796487,
          "hi": 0.7514176330865708
        },
        "ece": {
          "value": 0.03149522683147957,
          "lo": 0.026041321058523508,
          "hi": 0.041646159942767116
        },
        "reliability": {
          "value": 0.0016323515032443912,
          "lo": 0.0012245269658096858,
          "hi": 0.0029887046246071846
        },
        "resolution": {
          "value": 0.028251831702324568,
          "lo": 0.025786177624655605,
          "hi": 0.03159886042170951
        },
        "uncertainty": {
          "value": 0.19739215302103383,
          "lo": 0.19262450715283344,
          "hi": 0.20227536629943227
        },
        "auroc": {
          "value": 0.7446593400696117,
          "lo": 0.7326277586422778,
          "hi": 0.756000760333661
        },
        "acc@80": {
          "value": 0.7887129017376595,
          "lo": 0.7769549200023308,
          "hi": 0.8002794750066544
        }
      }
    },
    {
      "env": "chaosnli",
      "slice": "all",
      "system": "mix7-1.6B-s0",
      "metrics": {
        "n": {
          "value": 2763.0,
          "lo": 2732.475,
          "hi": 2798.525
        },
        "ce": {
          "value": 0.7654613715471141,
          "lo": 0.7494300351864432,
          "hi": 0.7808923184162763
        },
        "kl": {
          "value": 0.23510130147785244,
          "lo": 0.2222539775728993,
          "hi": 0.24790096655683252
        },
        "jsd": {
          "value": 0.05892347080822899,
          "lo": 0.05614685702665513,
          "hi": 0.06198860929577264
        },
        "brier": {
          "value": 0.1552037163302883,
          "lo": 0.1462400606765066,
          "hi": 0.16456635180066206
        },
        "acc": {
          "value": 0.7082880926529135,
          "lo": 0.6925717712394138,
          "hi": 0.7251270878721859
        },
        "conf": {
          "value": 0.6869863378065374,
          "lo": 0.681784769246876,
          "hi": 0.6923667213332382
        },
        "ece": {
          "value": 0.030578552813707745,
          "lo": 0.027840654001020076,
          "hi": 0.05437526135243692
        },
        "reliability": {
          "value": 0.0017135782171140309,
          "lo": 0.0012915285493501388,
          "hi": 0.004640477300022285
        },
        "resolution": {
          "value": 0.01976698573550043,
          "lo": 0.016656680968617162,
          "hi": 0.02524448990397694
        },
        "uncertainty": {
          "value": 0.20661607045901134,
          "lo": 0.19931779430618907,
          "hi": 0.21291611270607952
        },
        "auroc": {
          "value": 0.6902250748410934,
          "lo": 0.6678661067691528,
          "hi": 0.7123250156117648
        },
        "acc@80": {
          "value": 0.7529411764705882,
          "lo": 0.7348948919282172,
          "hi": 0.7722481802085382
        }
      }
    },
    {
      "env": "civilcomments",
      "slice": "all",
      "system": "mix7-1.6B-s0",
      "metrics": {
        "n": {
          "value": 10581.0,
          "lo": 10261.425,
          "hi": 10904.525
        },
        "ce": {
          "value": 0.3291432395939985,
          "lo": 0.3200820913634004,
          "hi": 0.3372185842634208
        },
        "kl": {
          "value": 0.09766043082465743,
          "lo": 0.09463633276139745,
          "hi": 0.10064454636867035
        },
        "jsd": {
          "value": 0.024450077575365353,
          "lo": 0.02376066723067054,
          "hi": 0.025111864127835334
        },
        "brier": {
          "value": 0.04824225276040297,
          "lo": 0.04608759857299234,
          "hi": 0.05039442895462397
        },
        "acc": {
          "value": 0.9314809564313392,
          "lo": 0.926637093159785,
          "hi": 0.9358851114843256
        },
        "conf": {
          "value": 0.8949943995861078,
          "lo": 0.8925153914028515,
          "hi": 0.8978972249872205
        },
        "ece": {
          "value": 0.04325364368555168,
          "lo": 0.039369716140716667,
          "hi": 0.04738038014011293
        },
        "reliability": {
          "value": 0.0025325939682709357,
          "lo": 0.0021693777868036232,
          "hi": 0.0031092331245068244
        },
        "resolution": {
          "value": 0.01635684924521661,
          "lo": 0.014259526254035334,
          "hi": 0.01878172823720776
        },
        "uncertainty": {
          "value": 0.06382418423709675,
          "lo": 0.06000416947908488,
          "hi": 0.06798078845455381
        },
        "auroc": {
          "value": 0.9217205693013882,
          "lo": 0.913332991673519,
          "hi": 0.9293901686812233
        },
        "acc@80": {
          "value": 0.9864146485528648,
          "lo": 0.9840827523103477,
          "hi": 0.9893232853040609
        }
      }
    },
    {
      "env": "folktables",
      "slice": "all",
      "system": "mix7-1.6B-s0",
      "metrics": {
        "n": {
          "value": 15003.0,
          "lo": 14855.8,
          "hi": 15157.0
        },
        "ce": {
          "value": 0.4429014798100703,
          "lo": 0.4338010939037464,
          "hi": 0.452151049808509
        },
        "kl": {
          "value": 0.4429014798100703,
          "lo": 0.4338010939037464,
          "hi": 0.452151049808509
        },
        "jsd": {
          "value": 0.1264743363940025,
          "lo": 0.12427204099846796,
          "hi": 0.12871952108589504
        },
        "brier": {
          "value": 0.29192483263490315,
          "lo": 0.28545596286425384,
          "hi": 0.2991082407855318
        },
        "acc": {
          "value": 0.7856428714257149,
          "lo": 0.7790224321133412,
          "hi": 0.7922019791889412
        },
        "conf": {
          "value": 0.8061644708856792,
          "lo": 0.8034163905686282,
          "hi": 0.8086048746732889
        },
        "ece": {
          "value": 0.025429919649106476,
          "lo": 0.020018190591396846,
          "hi": 0.03176325508409786
        },
        "reliability": {
          "value": 0.0009121385172829448,
          "lo": 0.0006445539894268695,
          "hi": 0.0015913915011417576
        },
        "resolution": {
          "value": 0.0233436276816435,
          "lo": 0.02173771272662814,
          "hi": 0.025429969496669848
        },
        "uncertainty": {
          "value": 0.1684081500036725,
          "lo": 0.16461800164228868,
          "hi": 0.1721464467827975
        },
        "auroc": {
          "value": 0.760233718887534,
          "lo": 0.7515386177960786,
          "hi": 0.7686806860142983
        },
        "acc@80": {
          "value": 0.8390268288618563,
          "lo": 0.831536650563158,
          "hi": 0.8459624211130234
        }
      }
    },
    {
      "env": "mcq",
      "slice": "all",
      "system": "mix7-1.6B-s0",
      "metrics": {
        "n": {
          "value": 5064.0,
          "lo": 5064.0,
          "hi": 5064.0
        },
        "ce": {
          "value": 0.4002523221645569,
          "lo": 0.3782136455299466,
          "hi": 0.42385603103379427
        },
        "kl": {
          "value": 0.4002523221645569,
          "lo": 0.3782136455299466,
          "hi": 0.42385603103379427
        },
        "jsd": {
          "value": 0.09391323294325246,
          "lo": 0.08972082333599198,
          "hi": 0.09825219031625738
        },
        "brier": {
          "value": 0.20633200495477827,
          "lo": 0.19470067976198033,
          "hi": 0.21846857959263977
        },
        "acc": {
          "value": 0.8554502369668247,
          "lo": 0.8450780015797789,
          "hi": 0.8641390205371248
        },
        "conf": {
          "value": 0.8843396415677963,
          "lo": 0.8795300210097814,
          "hi": 0.8890400585438909
        },
        "ece": {
          "value": 0.0288912705417964,
          "lo": 0.02272542280852676,
          "hi": 0.037411134562191764
        },
        "reliability": {
          "value": 0.0014989374336935519,
          "lo": 0.0010164991828886306,
          "hi": 0.002780364073085358
        },
        "resolution": {
          "value": 0.033634874124788244,
          "lo": 0.029467195088500375,
          "hi": 0.03864765960070014
        },
        "uncertainty": {
          "value": 0.1236551290402282,
          "lo": 0.11740277372226343,
          "hi": 0.1309211631012456
        },
        "auroc": {
          "value": 0.8781354225007443,
          "lo": 0.866011040404468,
          "hi": 0.8890746426752913
        },
        "acc@80": {
          "value": 0.9360651690940508,
          "lo": 0.9281658849666748,
          "hi": 0.9434707479634659
        }
      }
    },
    {
      "env": "stsb",
      "slice": "all",
      "system": "mix7-1.6B-s0",
      "metrics": {
        "n": {
          "value": 3676.0,
          "lo": 3150.95,
          "hi": 4793.674999999996
        },
        "ce": {
          "value": 1.031748552907778,
          "lo": 0.9704824263024805,
          "hi": 1.0889858187347534
        },
        "kl": {
          "value": 0.6294447300455221,
          "lo": 0.572015076742756,
          "hi": 0.6832772809458031
        },
        "jsd": {
          "value": 0.1479128524201686,
          "lo": 0.13660829407083266,
          "hi": 0.15784509493722043
        },
        "brier": {
          "value": 0.290818102811832,
          "lo": 0.2668247655785668,
          "hi": 0.3128366138275768
        },
        "acc": {
          "value": 0.6131664853101197,
          "lo": 0.587906790991269,
          "hi": 0.6378792112198279
        },
        "conf": {
          "value": 0.6504272062481422,
          "lo": 0.6443878978823263,
          "hi": 0.6562695090876604
        },
        "ece": {
          "value": 0.052070733139512326,
          "lo": 0.04119822485513943,
          "hi": 0.07324350903062836
        },
        "reliability": {
          "value": 0.004317376438774994,
          "lo": 0.0029592961485554003,
          "hi": 0.00820963383335681
        },
        "resolution": {
          "value": 0.013739284081678127,
          "lo": 0.011211957772358483,
          "hi": 0.017058838178443402
        },
        "uncertainty": {
          "value": 0.23719334660255445,
          "lo": 0.23098927933949753,
          "hi": 0.24227237669441876
        },
        "auroc": {
          "value": 0.6128823020677726,
          "lo": 0.5911372722639222,
          "hi": 0.6308539874239646
        },
        "acc@80": {
          "value": 0.6399183951037062,
          "lo": 0.6111209029066172,
          "hi": 0.6684116940898375
        }
      }
    },
    {
      "env": "yelp",
      "slice": "all",
      "system": "mix7-1.6B-s0",
      "metrics": {
        "n": {
          "value": 5000.0,
          "lo": 5000.0,
          "hi": 5000.0
        },
        "ce": {
          "value": 1.1007355521982354,
          "lo": 1.0810604797933898,
          "hi": 1.1197186777887949
        },
        "kl": {
          "value": 1.1007355521982354,
          "lo": 1.0810604797933898,
          "hi": 1.1197186777887949
        },
        "jsd": {
          "value": 0.2967980785000652,
          "lo": 0.2926581255188195,
          "hi": 0.3006860390537452
        },
        "brier": {
          "value": 0.6052061326700343,
          "lo": 0.5939172279744793,
          "hi": 0.6159469872070467
        },
        "acc": {
          "value": 0.4936,
          "lo": 0.4808,
          "hi": 0.507905
        },
        "conf": {
          "value": 0.5416603589355946,
          "lo": 0.5379687575377524,
          "hi": 0.5453135892054439
        },
        "ece": {
          "value": 0.055525891262292876,
          "lo": 0.04558760949969293,
          "hi": 0.06793060291126372
        },
        "reliability": {
          "value": 0.0037097829303959985,
          "lo": 0.0029671397564004914,
          "hi": 0.005947604738688515
        },
        "resolution": {
          "value": 0.021382927360414306,
          "lo": 0.0185277120501153,
          "hi": 0.025603394336124526
        },
        "uncertainty": {
          "value": 0.24995903999999997,
          "lo": 0.24963136,
          "hi": 0.24999996
        },
        "auroc": {
          "value": 0.6502098103753319,
          "lo": 0.6356542944329878,
          "hi": 0.6653851325094023
        },
        "acc@80": {
          "value": 0.5315,
          "lo": 0.5173687499999999,
          "hi": 0.548
        }
      }
    },
    {
      "env": "boolq",
      "slice": "all",
      "system": "mix7-1.6B-s1",
      "metrics": {
        "n": {
          "value": 8129.0,
          "lo": 7984.475,
          "hi": 8270.575
        },
        "ce": {
          "value": 0.47855710362970805,
          "lo": 0.4687038073149669,
          "hi": 0.48872637673033503
        },
        "kl": {
          "value": 0.47855710362970805,
          "lo": 0.4687038073149669,
          "hi": 0.48872637673033503
        },
        "jsd": {
          "value": 0.1457436762811214,
          "lo": 0.14307058632434067,
          "hi": 0.1485167780754198
        },
        "brier": {
          "value": 0.3130394035594896,
          "lo": 0.30479280208146486,
          "hi": 0.32176081123172
        },
        "acc": {
          "value": 0.7684832082666995,
          "lo": 0.7575482036495192,
          "hi": 0.7778365736030838
        },
        "conf": {
          "value": 0.7263465951562207,
          "lo": 0.7231753169988112,
          "hi": 0.7296469446485713
        },
        "ece": {
          "value": 0.04808772066331964,
          "lo": 0.04045296450417986,
          "hi": 0.058733391696430805
        },
        "reliability": {
          "value": 0.00284671903813426,
          "lo": 0.0022353916304481353,
          "hi": 0.004130891358852014
        },
        "resolution": {
          "value": 0.024018337935267288,
          "lo": 0.021604361539690864,
          "hi": 0.0270778336935212
        },
        "uncertainty": {
          "value": 0.17791676687882008,
          "lo": 0.17280683360798862,
          "hi": 0.18366891095742832
        },
        "auroc": {
          "value": 0.7496162663923529,
          "lo": 0.7382809384344194,
          "hi": 0.7626390674411822
        },
        "acc@80": {
          "value": 0.8236198677533446,
          "lo": 0.8130154888594103,
          "hi": 0.8347216561229348
        }
      }
    },
    {
      "env": "chaosnli",
      "slice": "all",
      "system": "mix7-1.6B-s1",
      "metrics": {
        "n": {
          "value": 2763.0,
          "lo": 2732.475,
          "hi": 2798.525
        },
        "ce": {
          "value": 0.7837159885599414,
          "lo": 0.7690258664923734,
          "hi": 0.7995680854416652
        },
        "kl": {
          "value": 0.25335591849067945,
          "lo": 0.24210794053634344,
          "hi": 0.26627656912373626
        },
        "jsd": {
          "value": 0.06430561168056584,
          "lo": 0.061732091418810385,
          "hi": 0.06738181425264714
        },
        "brier": {
          "value": 0.17074606097407263,
          "lo": 0.16227614013310068,
          "hi": 0.1806546230159187
        },
        "acc": {
          "value": 0.6800579080709374,
          "lo": 0.6616313268009419,
          "hi": 0.6974956007895736
        },
        "conf": {
          "value": 0.6740838724376929,
          "lo": 0.6684434502085462,
          "hi": 0.6796477558247118
        },
        "ece": {
          "value": 0.02181350362935101,
          "lo": 0.022559406720764993,
          "hi": 0.0476477255996329
        },
        "reliability": {
          "value": 0.0008491005387559562,
          "lo": 0.0008246849919371727,
          "hi": 0.0033061108419428875
        },
        "resolution": {
          "value": 0.017039685512144916,
          "lo": 0.01404967003529775,
          "hi": 0.02205710108209954
        },
        "uncertainty": {
          "value": 0.21757914974111786,
          "lo": 0.21099517127486303,
          "hi": 0.2238753103132754
        },
        "auroc": {
          "value": 0.6697049311393612,
          "lo": 0.6491910484953667,
          "hi": 0.6894518935356425
        },
        "acc@80": {
          "value": 0.7217194570135747,
          "lo": 0.7032542171221171,
          "hi": 0.7411502385227859
        }
      }
    },
    {
      "env": "civilcomments",
      "slice": "all",
      "system": "mix7-1.6B-s1",
      "metrics": {
        "n": {
          "value": 10581.0,
          "lo": 10261.425,
          "hi": 10904.525
        },
        "ce": {
          "value": 0.3231331506493501,
          "lo": 0.31510608432322923,
          "hi": 0.33053231851393033
        },
        "kl": {
          "value": 0.09165034188000905,
          "lo": 0.08921166635656741,
          "hi": 0.09423752245204604
        },
        "jsd": {
          "value": 0.02525167459760998,
          "lo": 0.02464559411887002,
          "hi": 0.025894354411833548
        },
        "brier": {
          "value": 0.04397100215369724,
          "lo": 0.04207733690750005,
          "hi": 0.04582011475573585
        },
        "acc": {
          "value": 0.93346564596919,
          "lo": 0.9287920101088212,
          "hi": 0.9380227139566573
        },
        "conf": {
          "value": 0.8726835281555306,
          "lo": 0.8698799446038222,
          "hi": 0.8757637342502513
        },
        "ece": {
          "value": 0.0636185640232424,
          "lo": 0.060011936270282715,
          "hi": 0.06774610592440335
        },
        "reliability": {
          "value": 0.005551196569429456,
          "lo": 0.0049828496460246514,
          "hi": 0.006304045412355839
        },
        "resolution": {
          "value": 0.01528220459349804,
          "lo": 0.01310670354760269,
          "hi": 0.017633533336201036
        },
        "uncertainty": {
          "value": 0.06210753376451282,
          "lo": 0.05813609646680648,
          "hi": 0.06613740925205552
        },
        "auroc": {
          "value": 0.9217670673143299,
          "lo": 0.9137446684305477,
          "hi": 0.9294152745063432
        },
        "acc@80": {
          "value": 0.9871234494979326,
          "lo": 0.9845517587908574,
          "hi": 0.9891930614162803
        }
      }
    },
    {
      "env": "folktables",
      "slice": "all",
      "system": "mix7-1.6B-s1",
      "metrics": {
        "n": {
          "value": 15003.0,
          "lo": 14855.8,
          "hi": 15157.0
        },
        "ce": {
          "value": 0.44795320351265,
          "lo": 0.43988770426826734,
          "hi": 0.45616144497070304
        },
        "kl": {
          "value": 0.44795320351265,
          "lo": 0.43988770426826734,
          "hi": 0.45616144497070304
        },
        "jsd": {
          "value": 0.1314335571367777,
          "lo": 0.12918742495131863,
          "hi": 0.13364927526286144
        },
        "brier": {
          "value": 0.29621629260680155,
          "lo": 0.28995125809083827,
          "hi": 0.3029004037214525
        },
        "acc": {
          "value": 0.7784443111377725,
          "lo": 0.7708365555562754,
          "hi": 0.7857097418027347
        },
        "conf": {
          "value": 0.7826558687977256,
          "lo": 0.7799437584074044,
          "hi": 0.7850506899493145
        },
        "ece": {
          "value": 0.013490881572951727,
          "lo": 0.010313860277834775,
          "hi": 0.02133988882347297
        },
        "reliability": {
          "value": 0.0002916299564866727,
          "lo": 0.00018029928017010852,
          "hi": 0.0007644078426685735
        },
        "resolution": {
          "value": 0.024531181398278194,
          "lo": 0.022702030203598764,
          "hi": 0.026773431321743524
        },
        "uncertainty": {
          "value": 0.17246876559501137,
          "lo": 0.16836993443494178,
          "hi": 0.17664756015099178
        },
        "auroc": {
          "value": 0.760897827041841,
          "lo": 0.7527185587025165,
          "hi": 0.7690317796217346
        },
        "acc@80": {
          "value": 0.8350274954174304,
          "lo": 0.8268750344847557,
          "hi": 0.8415655775646039
        }
      }
    },
    {
      "env": "mcq",
      "slice": "all",
      "system": "mix7-1.6B-s1",
      "metrics": {
        "n": {
          "value": 5064.0,
          "lo": 5064.0,
          "hi": 5064.0
        },
        "ce": {
          "value": 0.42684009817659985,
          "lo": 0.40534061865938326,
          "hi": 0.4487317974627279
        },
        "kl": {
          "value": 0.42684009817659985,
          "lo": 0.40534061865938326,
          "hi": 0.4487317974627279
        },
        "jsd": {
          "value": 0.1047209966086494,
          "lo": 0.10048340589225171,
          "hi": 0.10896929767667661
        },
        "brier": {
          "value": 0.21781855156334679,
          "lo": 0.2069730368823189,
          "hi": 0.22866386525479698
        },
        "acc": {
          "value": 0.846958925750395,
          "lo": 0.8374802527646129,
          "hi": 0.8564474723538705
        },
        "conf": {
          "value": 0.8572383006007703,
          "lo": 0.8524303276069745,
          "hi": 0.8613331286549292
        },
        "ece": {
          "value": 0.012092256853813047,
          "lo": 0.010227614650458664,
          "hi": 0.023183345897741164
        },
        "reliability": {
          "value": 0.00033246617151455745,
          "lo": 0.0002320958444557757,
          "hi": 0.001265645465417821
        },
        "resolution": {
          "value": 0.036602240533722234,
          "lo": 0.03218982097016048,
          "hi": 0.04174564202942582
        },
        "uncertainty": {
          "value": 0.1296195038421319,
          "lo": 0.12294516055469451,
          "hi": 0.13610707899393298
        },
        "auroc": {
          "value": 0.8766589700584391,
          "lo": 0.8656314490404682,
          "hi": 0.8881121225579844
        },
        "acc@80": {
          "value": 0.9313749691434213,
          "lo": 0.9230992347568501,
          "hi": 0.9381695877561097
        }
      }
    },
    {
      "env": "stsb",
      "slice": "all",
      "system": "mix7-1.6B-s1",
      "metrics": {
        "n": {
          "value": 3676.0,
          "lo": 3150.95,
          "hi": 4793.674999999996
        },
        "ce": {
          "value": 1.0325978650062761,
          "lo": 0.9812421226139131,
          "hi": 1.0807375838768523
        },
        "kl": {
          "value": 0.6302940421440205,
          "lo": 0.5805065812874415,
          "hi": 0.675634602555858
        },
        "jsd": {
          "value": 0.15664148876714856,
          "lo": 0.14586758188198548,
          "hi": 0.16587076106736395
        },
        "brier": {
          "value": 0.30203148903766214,
          "lo": 0.2793681708337265,
          "hi": 0.3214175970283663
        },
        "acc": {
          "value": 0.5715451577801959,
          "lo": 0.5432189681593734,
          "hi": 0.5975937224782812
        },
        "conf": {
          "value": 0.6035776622552866,
          "lo": 0.5966974025467761,
          "hi": 0.610775968141833
        },
        "ece": {
          "value": 0.04524248022022652,
          "lo": 0.03777464454166938,
          "hi": 0.07000553128061206
        },
        "reliability": {
          "value": 0.0029649726855611597,
          "lo": 0.002174174413508668,
          "hi": 0.0068987481306255465
        },
        "resolution": {
          "value": 0.013546995487004788,
          "lo": 0.010450661463123909,
          "hi": 0.01863973032653182
        },
        "uncertainty": {
          "value": 0.24488129039820689,
          "lo": 0.24047544437763765,
          "hi": 0.248132120610992
        },
        "auroc": {
          "value": 0.621027477467268,
          "lo": 0.5938392011645692,
          "hi": 0.6485523592260055
        },
        "acc@80": {
          "value": 0.6028561713702822,
          "lo": 0.5703271629887985,
          "hi": 0.6363633066398043
        }
      }
    },
    {
      "env": "yelp",
      "slice": "all",
      "system": "mix7-1.6B-s1",
      "metrics": {
        "n": {
          "value": 5000.0,
          "lo": 5000.0,
          "hi": 5000.0
        },
        "ce": {
          "value": 1.1098901509100594,
          "lo": 1.0902504703014342,
          "hi": 1.1298374929111197
        },
        "kl": {
          "value": 1.1098901509100594,
          "lo": 1.0902504703014342,
          "hi": 1.1298374929111197
        },
        "jsd": {
          "value": 0.29727174008141743,
          "lo": 0.29302459407091197,
          "hi": 0.3012404960787604
        },
        "brier": {
          "value": 0.5931743975488044,
          "lo": 0.582613042214466,
          "hi": 0.60352119345443
        },
        "acc": {
          "value": 0.5326,
          "lo": 0.518685,
          "hi": 0.5468
        },
        "conf": {
          "value": 0.5210890755355359,
          "lo": 0.5175688868150115,
          "hi": 0.5246254901348054
        },
        "ece": {
          "value": 0.031602025562524796,
          "lo": 0.027794011172354218,
          "hi": 0.049214339447766535
        },
        "reliability": {
          "value": 0.0015905999807410446,
          "lo": 0.0012336936696333237,
          "hi": 0.0034744437984095895
        },
        "resolution": {
          "value": 0.022269785719372066,
          "lo": 0.019280771406242864,
          "hi": 0.02686939814799849
        },
        "uncertainty": {
          "value": 0.24893724,
          "lo": 0.24780976000000002,
          "hi": 0.24965078100000002
        },
        "auroc": {
          "value": 0.6616861342240318,
          "lo": 0.6469168791201212,
          "hi": 0.677234165187451
        },
        "acc@80": {
          "value": 0.57175,
          "lo": 0.5550812500000001,
          "hi": 0.589
        }
      }
    },
    {
      "env": "boolq",
      "slice": "all",
      "system": "mix7-1.6B-s2",
      "metrics": {
        "n": {
          "value": 8129.0,
          "lo": 7984.475,
          "hi": 8270.575
        },
        "ce": {
          "value": 0.5184885789058133,
          "lo": 0.5073414469680039,
          "hi": 0.5299303650960395
        },
        "kl": {
          "value": 0.5184885789058133,
          "lo": 0.5073414469680039,
          "hi": 0.5299303650960395
        },
        "jsd": {
          "value": 0.15550945721477388,
          "lo": 0.1527205926314921,
          "hi": 0.15841855987947642
        },
        "brier": {
          "value": 0.3463320106654194,
          "lo": 0.33707788650964454,
          "hi": 0.35558891417168853
        },
        "acc": {
          "value": 0.7291179726903678,
          "lo": 0.7178177140234293,
          "hi": 0.7394308067209925
        },
        "conf": {
          "value": 0.7274356591758699,
          "lo": 0.7244902567629894,
          "hi": 0.7306223464699275
        },
        "ece": {
          "value": 0.03144833778070925,
          "lo": 0.026431504587073013,
          "hi": 0.04174819339554388
        },
        "reliability": {
          "value": 0.0014419736102994397,
          "lo": 0.0010707162982793146,
          "hi": 0.002594793016103829
        },
        "resolution": {
          "value": 0.025398090908325222,
          "lo": 0.022671965366050328,
          "hi": 0.028545983813892145
        },
        "uncertainty": {
          "value": 0.19750495459025585,
          "lo": 0.1926728720725454,
          "hi": 0.20255541854484804
        },
        "auroc": {
          "value": 0.7306054651913142,
          "lo": 0.7181527314846667,
          "hi": 0.7417360677234859
        },
        "acc@80": {
          "value": 0.7879440258342304,
          "lo": 0.7763449134687196,
          "hi": 0.798086545338904
        }
      }
    },
    {
      "env": "chaosnli",
      "slice": "all",
      "system": "mix7-1.6B-s2",
      "metrics": {
        "n": {
          "value": 2763.0,
          "lo": 2732.475,
          "hi": 2798.525
        },
        "ce": {
          "value": 0.7585695789984018,
          "lo": 0.7426975509629593,
          "hi": 0.7754835285504206
        },
        "kl": {
          "value": 0.22820950892914002,
          "lo": 0.21553206414863885,
          "hi": 0.24147519530133094
        },
        "jsd": {
          "value": 0.05680358621993229,
          "lo": 0.053984252957542944,
          "hi": 0.05979957873716277
        },
        "brier": {
          "value": 0.14927606706706414,
          "lo": 0.1402168151341663,
          "hi": 0.1586389948864686
        },
        "acc": {
          "value": 0.7278320665942816,
          "lo": 0.7109009257791143,
          "hi": 0.7449838551772038
        },
        "conf": {
          "value": 0.6846000137663043,
          "lo": 0.679424621606799,
          "hi": 0.6903244356008484
        },
        "ece": {
          "value": 0.056012450127855294,
          "lo": 0.04558204311582116,
          "hi": 0.07314523965095096
        },
        "reliability": {
          "value": 0.003940287369283433,
          "lo": 0.0031526820419978735,
          "hi": 0.007212572967384259
        },
        "resolution": {
          "value": 0.014707128696814295,
          "lo": 0.011672780792140328,
          "hi": 0.019684878450526758
        },
        "uncertainty": {
          "value": 0.19809254943137886,
          "lo": 0.1899829070334599,
          "hi": 0.20552079463651074
        },
        "auroc": {
          "value": 0.6720206417893078,
          "lo": 0.6508045463960422,
          "hi": 0.6950214079589969
        },
        "acc@80": {
          "value": 0.7737556561085973,
          "lo": 0.7553810211428797,
          "hi": 0.790371275790134
        }
      }
    },
    {
      "env": "civilcomments",
      "slice": "all",
      "system": "mix7-1.6B-s2",
      "metrics": {
        "n": {
          "value": 10581.0,
          "lo": 10261.425,
          "hi": 10904.525
        },
        "ce": {
          "value": 0.3163261854007762,
          "lo": 0.3083993356944391,
          "hi": 0.3236248116520279
        },
        "kl": {
          "value": 0.08484337663143518,
          "lo": 0.0824035445799588,
          "hi": 0.08768283157281037
        },
        "jsd": {
          "value": 0.022882750186549245,
          "lo": 0.022234285074514283,
          "hi": 0.0235707239570925
        },
        "brier": {
          "value": 0.039377961884232315,
          "lo": 0.03774909977347542,
          "hi": 0.04112938297858045
        },
        "acc": {
          "value": 0.9425385124279368,
          "lo": 0.9377526041043914,
          "hi": 0.9467558718433233
        },
        "conf": {
          "value": 0.8647348193914562,
          "lo": 0.8613458292080601,
          "hi": 0.8683777216504087
        },
        "ece": {
          "value": 0.07780369303648059,
          "lo": 0.07279760445520497,
          "hi": 0.08162182332821356
        },
        "reliability": {
          "value": 0.010334194443595839,
          "lo": 0.00929202710467513,
          "hi": 0.011360849811607052
        },
        "resolution": {
          "value": 0.01176697872619398,
          "lo": 0.010198096047943774,
          "hi": 0.014251711287922499
        },
        "uncertainty": {
          "value": 0.054159665018068785,
          "lo": 0.05040919081874604,
          "hi": 0.058372644645337014
        },
        "auroc": {
          "value": 0.9198111216072871,
          "lo": 0.9123547956416179,
          "hi": 0.927071480238284
        },
        "acc@80": {
          "value": 0.9890135853514471,
          "lo": 0.9865763365502729,
          "hi": 0.9912093413196587
        }
      }
    },
    {
      "env": "folktables",
      "slice": "all",
      "system": "mix7-1.6B-s2",
      "metrics": {
        "n": {
          "value": 15003.0,
          "lo": 14855.8,
          "hi": 15157.0
        },
        "ce": {
          "value": 0.4356590369299968,
          "lo": 0.4255747789775578,
          "hi": 0.44521351969369516
        },
        "kl": {
          "value": 0.4356590369299968,
          "lo": 0.4255747789775578,
          "hi": 0.44521351969369516
        },
        "jsd": {
          "value": 0.12351583298705354,
          "lo": 0.12126312193035323,
          "hi": 0.12589004464231174
        },
        "brier": {
          "value": 0.28450758265424486,
          "lo": 0.27780617593704,
          "hi": 0.29142304276175635
        },
        "acc": {
          "value": 0.7911084449776711,
          "lo": 0.784140465204479,
          "hi": 0.7979199537449984
        },
        "conf": {
          "value": 0.8088698498306295,
          "lo": 0.8060494976775232,
          "hi": 0.8115879565893355
        },
        "ece": {
          "value": 0.01979037306272671,
          "lo": 0.016355303584789916,
          "hi": 0.02714447983071726
        },
        "reliability": {
          "value": 0.000565135572096387,
          "lo": 0.00041583292822958926,
          "hi": 0.0010895546936945067
        },
        "resolution": {
          "value": 0.023620479731631654,
          "lo": 0.021853692097286808,
          "hi": 0.02577255861621752
        },
        "uncertainty": {
          "value": 0.16525587326268226,
          "lo": 0.16124369616200576,
          "hi": 0.16926419587388697
        },
        "auroc": {
          "value": 0.7657509738706254,
          "lo": 0.7575488803450572,
          "hi": 0.773991327057773
        },
        "acc@80": {
          "value": 0.8474420929845026,
          "lo": 0.840006326543138,
          "hi": 0.8532372466567733
        }
      }
    },
    {
      "env": "mcq",
      "slice": "all",
      "system": "mix7-1.6B-s2",
      "metrics": {
        "n": {
          "value": 5064.0,
          "lo": 5064.0,
          "hi": 5064.0
        },
        "ce": {
          "value": 0.41635328038970587,
          "lo": 0.39123240415937593,
          "hi": 0.44061227208142756
        },
        "kl": {
          "value": 0.41635328038970587,
          "lo": 0.39123240415937593,
          "hi": 0.44061227208142756
        },
        "jsd": {
          "value": 0.09177637644129429,
          "lo": 0.0874378979350158,
          "hi": 0.09609393418574062
        },
        "brier": {
          "value": 0.2106166671020026,
          "lo": 0.19884213068170087,
          "hi": 0.22260596854163442
        },
        "acc": {
          "value": 0.8578199052132701,
          "lo": 0.8488299763033176,
          "hi": 0.8672146524486571
        },
        "conf": {
          "value": 0.9008063965758064,
          "lo": 0.8966689786369108,
          "hi": 0.904753290160337
        },
        "ece": {
          "value": 0.0429864913625363,
          "lo": 0.036125414531183606,
          "hi": 0.051237418890399566
        },
        "reliability": {
          "value": 0.00413420947410155,
          "lo": 0.0028608661129630114,
          "hi": 0.005898513586923716
        },
        "resolution": {
          "value": 0.032241344066748535,
          "lo": 0.027679393936887287,
          "hi": 0.03666930523992034
        },
        "uncertainty": {
          "value": 0.12196491543316638,
          "lo": 0.11515331150686695,
          "hi": 0.12831763790776637
        },
        "auroc": {
          "value": 0.8752545017393084,
          "lo": 0.8638134438304621,
          "hi": 0.8857928861927227
        },
        "acc@80": {
          "value": 0.9333497901752653,
          "lo": 0.9259442113058504,
          "hi": 0.9411318193038757
        }
      }
    },
    {
      "env": "stsb",
      "slice": "all",
      "system": "mix7-1.6B-s2",
      "metrics": {
        "n": {
          "value": 3676.0,
          "lo": 3150.95,
          "hi": 4793.674999999996
        },
        "ce": {
          "value": 1.1350048984988437,
          "lo": 1.059055504232052,
          "hi": 1.2076017137835424
        },
        "kl": {
          "value": 0.732701075636588,
          "lo": 0.6608848588343845,
          "hi": 0.8045771003578774
        },
        "jsd": {
          "value": 0.1485437192618815,
          "lo": 0.1363266804954933,
          "hi": 0.15994733514458628
        },
        "brier": {
          "value": 0.3191667956646663,
          "lo": 0.29302127439037406,
          "hi": 0.34358278105281215
        },
        "acc": {
          "value": 0.5998367791077258,
          "lo": 0.5756276123301985,
          "hi": 0.6217297290540421
        },
        "conf": {
          "value": 0.721857658172912,
          "lo": 0.7139604273591086,
          "hi": 0.7298799076292077
        },
        "ece": {
          "value": 0.12202087906518622,
          "lo": 0.10624456008692325,
          "hi": 0.14103889926102078
        },
        "reliability": {
          "value": 0.01820899459314394,
          "lo": 0.015093344534630931,
          "hi": 0.02479185368774445
        },
        "resolution": {
          "value": 0.015581056345493581,
          "lo": 0.011424916235943822,
          "hi": 0.021824644160194848
        },
        "uncertainty": {
          "value": 0.24003261753739516,
          "lo": 0.23518186354343104,
          "hi": 0.2442793155504878
        },
        "auroc": {
          "value": 0.6290156941997284,
          "lo": 0.600513688902862,
          "hi": 0.6555364844263627
        },
        "acc@80": {
          "value": 0.6290377422645359,
          "lo": 0.5976708068202223,
          "hi": 0.6559320254312636
        }
      }
    },
    {
      "env": "yelp",
      "slice": "all",
      "system": "mix7-1.6B-s2",
      "metrics": {
        "n": {
          "value": 5000.0,
          "lo": 5000.0,
          "hi": 5000.0
        },
        "ce": {
          "value": 1.1317714985484708,
          "lo": 1.1070318971483937,
          "hi": 1.1563624192840993
        },
        "kl": {
          "value": 1.1317714985484708,
          "lo": 1.1070318971483937,
          "hi": 1.1563624192840993
        },
        "jsd": {
          "value": 0.28937557216683035,
          "lo": 0.28447826678600413,
          "hi": 0.29423930364960643
        },
        "brier": {
          "value": 0.6464389177537861,
          "lo": 0.6331029659381819,
          "hi": 0.6603220935009174
        },
        "acc": {
          "value": 0.484,
          "lo": 0.4708,
          "hi": 0.497305
        },
        "conf": {
          "value": 0.6399741260707379,
          "lo": 0.6360008893069625,
          "hi": 0.6436980085825921
        },
        "ece": {
          "value": 0.15597412607073785,
          "lo": 0.14327579551830888,
          "hi": 0.168145623575598
        },
        "reliability": {
          "value": 0.026895021035463172,
          "lo": 0.02383132064565059,
          "hi": 0.03218924544504216
        },
        "resolution": {
          "value": 0.014863111326895757,
          "lo": 0.012713097310915848,
          "hi": 0.01878409197471433
        },
        "uncertainty": {
          "value": 0.249744,
          "lo": 0.24914736,
          "hi": 0.24999216
        },
        "auroc": {
          "value": 0.628967823050804,
          "lo": 0.6138308150497146,
          "hi": 0.6436885618805029
        },
        "acc@80": {
          "value": 0.519,
          "lo": 0.5033687499999999,
          "hi": 0.5341312500000001
        }
      }
    },
    {
      "env": "boolq",
      "slice": "all",
      "system": "mix7-450M-s0",
      "metrics": {
        "n": {
          "value": 8129.0,
          "lo": 7984.475,
          "hi": 8270.575
        },
        "ce": {
          "value": 0.7136019924352867,
          "lo": 0.6992330314186738,
          "hi": 0.7284831141138027
        },
        "kl": {
          "value": 0.7136019924352867,
          "lo": 0.6992330314186738,
          "hi": 0.7284831141138027
        },
        "jsd": {
          "value": 0.1997060328901781,
          "lo": 0.19629395072043002,
          "hi": 0.2034049058664722
        },
        "brier": {
          "value": 0.50396008392508,
          "lo": 0.4924207562288041,
          "hi": 0.516100548797697
        },
        "acc": {
          "value": 0.5961372862590725,
          "lo": 0.5833206151993764,
          "hi": 0.6072701905987213
        },
        "conf": {
          "value": 0.7473454058632256,
          "lo": 0.7440006008262551,
          "hi": 0.750466218302769
        },
        "ece": {
          "value": 0.15120811960415317,
          "lo": 0.14059334256144637,
          "hi": 0.1628602029115415
        },
        "reliability": {
          "value": 0.025790954004519116,
          "lo": 0.022849653178952548,
          "hi": 0.030032748048394644
        },
        "resolution": {
          "value": 0.014443398893949458,
          "lo": 0.012634056254531721,
          "hi": 0.01717559926160552
        },
        "uncertainty": {
          "value": 0.24075762219074115,
          "lo": 0.23849310329488035,
          "hi": 0.24305724425008088
        },
        "auroc": {
          "value": 0.6325759999517268,
          "lo": 0.6216105301613887,
          "hi": 0.6446417077529917
        },
        "acc@80": {
          "value": 0.6287867138243888,
          "lo": 0.6141491061740227,
          "hi": 0.6399709257842388
        }
      }
    },
    {
      "env": "chaosnli",
      "slice": "all",
      "system": "mix7-450M-s0",
      "metrics": {
        "n": {
          "value": 2763.0,
          "lo": 2732.475,
          "hi": 2798.525
        },
        "ce": {
          "value": 0.8982318895558363,
          "lo": 0.8809683495606556,
          "hi": 0.9142790524229913
        },
        "kl": {
          "value": 0.3678718194865745,
          "lo": 0.35238190600853664,
          "hi": 0.384989532036147
        },
        "jsd": {
          "value": 0.09134736494105902,
          "lo": 0.08804191118690168,
          "hi": 0.09513120937645976
        },
        "brier": {
          "value": 0.24486443443340125,
          "lo": 0.23463086135651814,
          "hi": 0.25658701777405735
        },
        "acc": {
          "value": 0.5696706478465436,
          "lo": 0.5515324548447733,
          "hi": 0.5888673090601938
        },
        "conf": {
          "value": 0.6321799852064542,
          "lo": 0.627721041997912,
          "hi": 0.6365111712155478
        },
        "ece": {
          "value": 0.07636093434284788,
          "lo": 0.06381985569042921,
          "hi": 0.09822724828004374
        },
        "reliability": {
          "value": 0.008408304299897924,
          "lo": 0.006805944793064514,
          "hi": 0.014170574047492654
        },
        "resolution": {
          "value": 0.00602359624024682,
          "lo": 0.004797682776253241,
          "hi": 0.01106953877313359
        },
        "uncertainty": {
          "value": 0.2451460008286429,
          "lo": 0.24210259247819907,
          "hi": 0.24734439344902914
        },
        "auroc": {
          "value": 0.5786139997841288,
          "lo": 0.5569645510149233,
          "hi": 0.6020625222577087
        },
        "acc@80": {
          "value": 0.5923076923076923,
          "lo": 0.5700111036152837,
          "hi": 0.6124531330235895
        }
      }
    },
    {
      "env": "civilcomments",
      "slice": "all",
      "system": "mix7-450M-s0",
      "metrics": {
        "n": {
          "value": 10581.0,
          "lo": 10261.425,
          "hi": 10904.525
        },
        "ce": {
          "value": 0.3362536541996374,
          "lo": 0.32705962616964074,
          "hi": 0.3443111084010556
        },
        "kl": {
          "value": 0.1047708454302963,
          "lo": 0.10134550620320375,
          "hi": 0.10809404029584808
        },
        "jsd": {
          "value": 0.0258384918133153,
          "lo": 0.025078699358151307,
          "hi": 0.02657033880076522
        },
        "brier": {
          "value": 0.05199782543309763,
          "lo": 0.04963794515746586,
          "hi": 0.05439872462029978
        },
        "acc": {
          "value": 0.9291182307910405,
          "lo": 0.9243398642858486,
          "hi": 0.933591687041009
        },
        "conf": {
          "value": 0.8974095615014813,
          "lo": 0.8949546016420573,
          "hi": 0.9002264280977454
        },
        "ece": {
          "value": 0.03717635887555369,
          "lo": 0.033073985492604104,
          "hi": 0.04094511516000459
        },
        "reliability": {
          "value": 0.0018100862451617744,
          "lo": 0.0015027982301821941,
          "hi": 0.002249096987843986
        },
        "resolution": {
          "value": 0.01567649580728653,
          "lo": 0.013568188613809378,
          "hi": 0.01819872012798344
        },
        "uncertainty": {
          "value": 0.0658575440027673,
          "lo": 0.06199824016764933,
          "hi": 0.06993567843806958
        },
        "auroc": {
          "value": 0.9130753060048147,
          "lo": 0.904983118439983,
          "hi": 0.9213973689142747
        },
        "acc@80": {
          "value": 0.9845245126993503,
          "lo": 0.9811782355859234,
          "hi": 0.9873262110974187
        }
      }
    },
    {
      "env": "folktables",
      "slice": "all",
      "system": "mix7-450M-s0",
      "metrics": {
        "n": {
          "value": 15003.0,
          "lo": 14855.8,
          "hi": 15157.0
        },
        "ce": {
          "value": 0.43932493729908706,
          "lo": 0.4306063873085365,
          "hi": 0.44795527539062224
        },
        "kl": {
          "value": 0.43932493729908706,
          "lo": 0.4306063873085365,
          "hi": 0.44795527539062224
        },
        "jsd": {
          "value": 0.12698998066490708,
          "lo": 0.1247512801296757,
          "hi": 0.1291667144425798
        },
        "brier": {
          "value": 0.28949881585732284,
          "lo": 0.28289448648859516,
          "hi": 0.2962732345100782
        },
        "acc": {
          "value": 0.7843098047057255,
          "lo": 0.7771967479746235,
          "hi": 0.7909794763309138
        },
        "conf": {
          "value": 0.7981700179879287,
          "lo": 0.795316481521177,
          "hi": 0.8008636433840136
        },
        "ece": {
          "value": 0.016622621027984567,
          "lo": 0.012594631006075246,
          "hi": 0.023673511842104707
        },
        "reliability": {
          "value": 0.00041057214653177586,
          "lo": 0.00028476222813153264,
          "hi": 0.0009173206104042607
        },
        "resolution": {
          "value": 0.024776907383988914,
          "lo": 0.022891319809217686,
          "hi": 0.02678481059917013
        },
        "uncertainty": {
          "value": 0.1691679349481922,
          "lo": 0.16533094318620042,
          "hi": 0.1731619621170642
        },
        "auroc": {
          "value": 0.7666138137673784,
          "lo": 0.7587729096798848,
          "hi": 0.7744423019809946
        },
        "acc@80": {
          "value": 0.8434427595400766,
          "lo": 0.8364590299690824,
          "hi": 0.8505335929509873
        }
      }
    },
    {
      "env": "mcq",
      "slice": "all",
      "system": "mix7-450M-s0",
      "metrics": {
        "n": {
          "value": 5064.0,
          "lo": 5064.0,
          "hi": 5064.0
        },
        "ce": {
          "value": 0.747297009316037,
          "lo": 0.7213950937827313,
          "hi": 0.7766705758301444
        },
        "kl": {
          "value": 0.747297009316037,
          "lo": 0.7213950937827313,
          "hi": 0.7766705758301444
        },
        "jsd": {
          "value": 0.1767239562795029,
          "lo": 0.17163378414726665,
          "hi": 0.18206793350034872
        },
        "brier": {
          "value": 0.38947030443786845,
          "lo": 0.37626890786001316,
          "hi": 0.4035176613209252
        },
        "acc": {
          "value": 0.7156398104265402,
          "lo": 0.7043838862559242,
          "hi": 0.7273943522906793
        },
        "conf": {
          "value": 0.7855074125037516,
          "lo": 0.7796935032984745,
          "hi": 0.7909096821161254
        },
        "ece": {
          "value": 0.06986760207721124,
          "lo": 0.060243255898023915,
          "hi": 0.08024064675863614
        },
        "reliability": {
          "value": 0.006471904151593135,
          "lo": 0.005258512902660889,
          "hi": 0.00894751923819154
        },
        "resolution": {
          "value": 0.048900232407484534,
          "lo": 0.04468738679701923,
          "hi": 0.05468343522446478
        },
        "uncertainty": {
          "value": 0.2034994721592058,
          "lo": 0.19829179882184186,
          "hi": 0.20822722703892543
        },
        "auroc": {
          "value": 0.8112850096578367,
          "lo": 0.7978697372086834,
          "hi": 0.8238597453353185
        },
        "acc@80": {
          "value": 0.7965934337200691,
          "lo": 0.7846149098987903,
          "hi": 0.8086892125401135
        }
      }
    },
    {
      "env": "stsb",
      "slice": "all",
      "system": "mix7-450M-s0",
      "metrics": {
        "n": {
          "value": 3676.0,
          "lo": 3150.95,
          "hi": 4793.674999999996
        },
        "ce": {
          "value": 1.1119125656921782,
          "lo": 1.053280421966601,
          "hi": 1.1685552061846858
        },
        "kl": {
          "value": 0.7096087428299227,
          "lo": 0.65358620171179,
          "hi": 0.766083341781686
        },
        "jsd": {
          "value": 0.17320132983767303,
          "lo": 0.1616274604120623,
          "hi": 0.18393198896815666
        },
        "brier": {
          "value": 0.3273993070991984,
          "lo": 0.3053723911340578,
          "hi": 0.3496214053801269
        },
        "acc": {
          "value": 0.5525027203482046,
          "lo": 0.5235303024329845,
          "hi": 0.5782422389407145
        },
        "conf": {
          "value": 0.5877989736487997,
          "lo": 0.5764306252561174,
          "hi": 0.6025828143294766
        },
        "ece": {
          "value": 0.04876845997228197,
          "lo": 0.03908335157774449,
          "hi": 0.06955720642802253
        },
        "reliability": {
          "value": 0.0029641416367139854,
          "lo": 0.002345042505377347,
          "hi": 0.006676650159377232
        },
        "resolution": {
          "value": 0.013711604731963535,
          "lo": 0.011535638417900448,
          "hi": 0.01873196110357933
        },
        "uncertainty": {
          "value": 0.24724346435603822,
          "lo": 0.2438781452883429,
          "hi": 0.24944630452685188
        },
        "auroc": {
          "value": 0.6062123708655655,
          "lo": 0.58165551110918,
          "hi": 0.6276101089025438
        },
        "acc@80": {
          "value": 0.5725943556613396,
          "lo": 0.5416357437692293,
          "hi": 0.6052191508252037
        }
      }
    },
    {
      "env": "yelp",
      "slice": "all",
      "system": "mix7-450M-s0",
      "metrics": {
        "n": {
          "value": 5000.0,
          "lo": 5000.0,
          "hi": 5000.0
        },
        "ce": {
          "value": 1.4732754178012324,
          "lo": 1.4513124485509135,
          "hi": 1.4966784315799533
        },
        "kl": {
          "value": 1.4732754178012324,
          "lo": 1.4513124485509135,
          "hi": 1.4966784315799533
        },
        "jsd": {
          "value": 0.3673805956929661,
          "lo": 0.362933915344585,
          "hi": 0.37171122393183115
        },
        "brier": {
          "value": 0.7798598577553912,
          "lo": 0.767312370573482,
          "hi": 0.7911860416874497
        },
        "acc": {
          "value": 0.3442,
          "lo": 0.3314,
          "hi": 0.35841
        },
        "conf": {
          "value": 0.5286268295913935,
          "lo": 0.5250381398198753,
          "hi": 0.532158047472313
        },
        "ece": {
          "value": 0.18442682959139348,
          "lo": 0.17059637910090386,
          "hi": 0.1975069882760942
        },
        "reliability": {
          "value": 0.04259200737759276,
          "lo": 0.03767236290099163,
          "hi": 0.04908729373114615
        },
        "resolution": {
          "value": 0.0042384781061300825,
          "lo": 0.0033056711060300992,
          "hi": 0.006562855173325714
        },
        "uncertainty": {
          "value": 0.22572636,
          "lo": 0.22157404,
          "hi": 0.22995223199999998
        },
        "auroc": {
          "value": 0.5572634051246829,
          "lo": 0.5406379019247921,
          "hi": 0.5753965855237008
        },
        "acc@80": {
          "value": 0.36725,
          "lo": 0.3512125,
          "hi": 0.3832625
        }
      }
    },
    {
      "env": "boolq",
      "slice": "all",
      "system": "mix7-450M-s1",
      "metrics": {
        "n": {
          "value": 8129.0,
          "lo": 7984.475,
          "hi": 8270.575
        },
        "ce": {
          "value": 0.6613666478181621,
          "lo": 0.6487013682148409,
          "hi": 0.6745199642192672
        },
        "kl": {
          "value": 0.6613666478181621,
          "lo": 0.6487013682148409,
          "hi": 0.6745199642192672
        },
        "jsd": {
          "value": 0.19066056494487488,
          "lo": 0.18749099463660807,
          "hi": 0.19401811688218176
        },
        "brier": {
          "value": 0.4651263812087598,
          "lo": 0.45469805941530567,
          "hi": 0.4759976378084498
        },
        "acc": {
          "value": 0.6204945257719277,
          "lo": 0.6092188247359599,
          "hi": 0.6309430661660926
        },
        "conf": {
          "value": 0.7238843866872207,
          "lo": 0.7205719296032946,
          "hi": 0.7267868473831358
        },
        "ece": {
          "value": 0.10346502359685661,
          "lo": 0.09424817112916596,
          "hi": 0.11470731441404307
        },
        "reliability": {
          "value": 0.0130333503843448,
          "lo": 0.01105717569199151,
          "hi": 0.01607895916415323
        },
        "resolution": {
          "value": 0.015713242092913044,
          "lo": 0.013837220264048737,
          "hi": 0.0182541093085967
        },
        "uncertainty": {
          "value": 0.23548106925899825,
          "lo": 0.2328539124458066,
          "hi": 0.23807119371158414
        },
        "auroc": {
          "value": 0.6412828374486046,
          "lo": 0.6283801720039986,
          "hi": 0.6536368927242742
        },
        "acc@80": {
          "value": 0.6513916653852069,
          "lo": 0.6387216888372156,
          "hi": 0.6638142819075533
        }
      }
    },
    {
      "env": "chaosnli",
      "slice": "all",
      "system": "mix7-450M-s1",
      "metrics": {
        "n": {
          "value": 2763.0,
          "lo": 2732.475,
          "hi": 2798.525
        },
        "ce": {
          "value": 0.9017071316185276,
          "lo": 0.8833701893435971,
          "hi": 0.9209804894390085
        },
        "kl": {
          "value": 0.37134706154926567,
          "lo": 0.3545081413060039,
          "hi": 0.39115978645605487
        },
        "jsd": {
          "value": 0.09162369502806199,
          "lo": 0.08801208138497119,
          "hi": 0.09608325834988683
        },
        "brier": {
          "value": 0.24518685638804466,
          "lo": 0.23340681051390524,
          "hi": 0.25813472844184493
        },
        "acc": {
          "value": 0.5754614549402823,
          "lo": 0.5559670321869261,
          "hi": 0.5955889039843612
        },
        "conf": {
          "value": 0.6433491620774119,
          "lo": 0.6386417505573803,
          "hi": 0.6481729690973095
        },
        "ece": {
          "value": 0.07894042577598563,
          "lo": 0.06663756892339306,
          "hi": 0.09771193973001406
        },
        "reliability": {
          "value": 0.010354522866363586,
          "lo": 0.008124692146806126,
          "hi": 0.015591151725612823
        },
        "resolution": {
          "value": 0.00784090507888757,
          "lo": 0.005835860267174734,
          "hi": 0.012494261150418014
        },
        "uncertainty": {
          "value": 0.24430556881829574,
          "lo": 0.24086262703703576,
          "hi": 0.2468676430223986
        },
        "auroc": {
          "value": 0.5909461843255213,
          "lo": 0.5695110838859726,
          "hi": 0.6109242883847428
        },
        "acc@80": {
          "value": 0.5972850678733032,
          "lo": 0.5770786718246081,
          "hi": 0.6213368257061813
        }
      }
    },
    {
      "env": "civilcomments",
      "slice": "all",
      "system": "mix7-450M-s1",
      "metrics": {
        "n": {
          "value": 10581.0,
          "lo": 10261.425,
          "hi": 10904.525
        },
        "ce": {
          "value": 0.32425900287305687,
          "lo": 0.3163181434463966,
          "hi": 0.33175254743750576
        },
        "kl": {
          "value": 0.0927761941037158,
          "lo": 0.0899606203264132,
          "hi": 0.09570525052304658
        },
        "jsd": {
          "value": 0.02541242196432821,
          "lo": 0.02467779413471958,
          "hi": 0.026152304085951525
        },
        "brier": {
          "value": 0.044416175920525686,
          "lo": 0.04240057062755955,
          "hi": 0.04647481358668125
        },
        "acc": {
          "value": 0.9396087326339665,
          "lo": 0.9352888967187092,
          "hi": 0.9439752403673666
        },
        "conf": {
          "value": 0.8663100878296226,
          "lo": 0.8631051099045708,
          "hi": 0.8698803377815415
        },
        "ece": {
          "value": 0.07329864480434391,
          "lo": 0.06895776154217659,
          "hi": 0.07696218223674356
        },
        "reliability": {
          "value": 0.007928800038177412,
          "lo": 0.0070054302122457,
          "hi": 0.00882283596655696
        },
        "resolution": {
          "value": 0.011163723718570301,
          "lo": 0.009838755300416056,
          "hi": 0.012906397767968134
        },
        "uncertainty": {
          "value": 0.05674416219195776,
          "lo": 0.05288598575935794,
          "hi": 0.06052357633528109
        },
        "auroc": {
          "value": 0.9061359484383446,
          "lo": 0.8971158509009561,
          "hi": 0.9145749425838893
        },
        "acc@80": {
          "value": 0.9866509155345541,
          "lo": 0.9836346548870232,
          "hi": 0.9891896272006314
        }
      }
    },
    {
      "env": "folktables",
      "slice": "all",
      "system": "mix7-450M-s1",
      "metrics": {
        "n": {
          "value": 15003.0,
          "lo": 14855.8,
          "hi": 15157.0
        },
        "ce": {
          "value": 0.4356261342600311,
          "lo": 0.4262985371348895,
          "hi": 0.44383308741446337
        },
        "kl": {
          "value": 0.4356261342600311,
          "lo": 0.4262985371348895,
          "hi": 0.44383308741446337
        },
        "jsd": {
          "value": 0.12577303288797503,
          "lo": 0.12356107449550632,
          "hi": 0.12789032184153792
        },
        "brier": {
          "value": 0.2863458999243068,
          "lo": 0.2798399479718523,
          "hi": 0.29287672179440166
        },
        "acc": {
          "value": 0.789375458241685,
          "lo": 0.782430143058245,
          "hi": 0.7961437083938346
        },
        "conf": {
          "value": 0.7976119624434539,
          "lo": 0.7947091525830844,
          "hi": 0.8002880792451132
        },
        "ece": {
          "value": 0.013219735506875569,
          "lo": 0.011337052212381292,
          "hi": 0.020845035677029018
        },
        "reliability": {
          "value": 0.0003257695476476921,
          "lo": 0.00024292459016608795,
          "hi": 0.0007882323891700785
        },
        "resolution": {
          "value": 0.02332887340217792,
          "lo": 0.02159768656050772,
          "hi": 0.0255286686995298
        },
        "uncertainty": {
          "value": 0.16626184416741482,
          "lo": 0.16229890348613865,
          "hi": 0.17023321077601278
        },
        "auroc": {
          "value": 0.7633103515723115,
          "lo": 0.7554848240124551,
          "hi": 0.7721751096809657
        },
        "acc@80": {
          "value": 0.8421929678386936,
          "lo": 0.8359579924822174,
          "hi": 0.848862181247759
        }
      }
    },
    {
      "env": "mcq",
      "slice": "all",
      "system": "mix7-450M-s1",
      "metrics": {
        "n": {
          "value": 5064.0,
          "lo": 5064.0,
          "hi": 5064.0
        },
        "ce": {
          "value": 0.7268390369943714,
          "lo": 0.701661708648479,
          "hi": 0.7544727097001674
        },
        "kl": {
          "value": 0.7268390369943714,
          "lo": 0.701661708648479,
          "hi": 0.7544727097001674
        },
        "jsd": {
          "value": 0.1737407686543726,
          "lo": 0.1686769082277473,
          "hi": 0.17902775793235276
        },
        "brier": {
          "value": 0.38048971460670034,
          "lo": 0.36697716722425716,
          "hi": 0.39513580937699744
        },
        "acc": {
          "value": 0.7215639810426541,
          "lo": 0.7087282780410743,
          "hi": 0.733214849921011
        },
        "conf": {
          "value": 0.7886344792283071,
          "lo": 0.7825879776870941,
          "hi": 0.7938439071054848
        },
        "ece": {
          "value": 0.06707049818565308,
          "lo": 0.05728493071724713,
          "hi": 0.07796081099659204
        },
        "reliability": {
          "value": 0.006713542783910689,
          "lo": 0.005178319951393914,
          "hi": 0.00933968993179277
        },
        "resolution": {
          "value": 0.05029871507191342,
          "lo": 0.04609185458371129,
          "hi": 0.05585772583013485
        },
        "uncertainty": {
          "value": 0.20090940230453042,
          "lo": 0.1956108337763203,
          "hi": 0.20643250594600798
        },
        "auroc": {
          "value": 0.8156515544996836,
          "lo": 0.8048366961594599,
          "hi": 0.8276977452680627
        },
        "acc@80": {
          "value": 0.800049370525796,
          "lo": 0.7878116514440878,
          "hi": 0.8131449024932116
        }
      }
    },
    {
      "env": "stsb",
      "slice": "all",
      "system": "mix7-450M-s1",
      "metrics": {
        "n": {
          "value": 3676.0,
          "lo": 3150.95,
          "hi": 4793.674999999996
        },
        "ce": {
          "value": 1.1069691649903395,
          "lo": 1.0446012821079593,
          "hi": 1.165738121532312
        },
        "kl": {
          "value": 0.7046653421280841,
          "lo": 0.6449676716730225,
          "hi": 0.760569363090012
        },
        "jsd": {
          "value": 0.16901927548315193,
          "lo": 0.1571037350910772,
          "hi": 0.17973577357612286
        },
        "brier": {
          "value": 0.32481970890073175,
          "lo": 0.3027704832669146,
          "hi": 0.3472728203806762
        },
        "acc": {
          "value": 0.5620239390642002,
          "lo": 0.5360869489613143,
          "hi": 0.5870780451619467
        },
        "conf": {
          "value": 0.6037559004194459,
          "lo": 0.5937292598986336,
          "hi": 0.6165383115026002
        },
        "ece": {
          "value": 0.04551832948662383,
          "lo": 0.036274910117460646,
          "hi": 0.06540883176745567
        },
        "reliability": {
          "value": 0.0029292827612687512,
          "lo": 0.0019677464215818396,
          "hi": 0.005977527992258667
        },
        "resolution": {
          "value": 0.013574186923772965,
          "lo": 0.010217866698878853,
          "hi": 0.017759200429510277
        },
        "uncertainty": {
          "value": 0.24615303098296037,
          "lo": 0.2424174069956571,
          "hi": 0.24869771644786673
        },
        "auroc": {
          "value": 0.6223150024351675,
          "lo": 0.6008536046440507,
          "hi": 0.640362722022482
        },
        "acc@80": {
          "value": 0.5946956817409045,
          "lo": 0.5682574797266854,
          "hi": 0.6205814035115087
        }
      }
    },
    {
      "env": "yelp",
      "slice": "all",
      "system": "mix7-450M-s1",
      "metrics": {
        "n": {
          "value": 5000.0,
          "lo": 5000.0,
          "hi": 5000.0
        },
        "ce": {
          "value": 1.4667452588920318,
          "lo": 1.444240469510126,
          "hi": 1.49047082016875
        },
        "kl": {
          "value": 1.4667452588920318,
          "lo": 1.444240469510126,
          "hi": 1.49047082016875
        },
        "jsd": {
          "value": 0.36672095432867907,
          "lo": 0.36240916995821554,
          "hi": 0.3710108191970698
        },
        "brier": {
          "value": 0.7609335954239111,
          "lo": 0.7495009082898206,
          "hi": 0.7717066495198419
        },
        "acc": {
          "value": 0.3754,
          "lo": 0.36138,
          "hi": 0.3894
        },
        "conf": {
          "value": 0.5026116242468357,
          "lo": 0.499141373424083,
          "hi": 0.5060167957246303
        },
        "ece": {
          "value": 0.1303246833860874,
          "lo": 0.11905975172311067,
          "hi": 0.1465369994214177
        },
        "reliability": {
          "value": 0.0262975168219169,
          "lo": 0.022168531149260436,
          "hi": 0.031952602452765634
        },
        "resolution": {
          "value": 0.0025656878872772973,
          "lo": 0.002004726317232737,
          "hi": 0.004824254096524662
        },
        "uncertainty": {
          "value": 0.23447484000000002,
          "lo": 0.230784336,
          "hi": 0.23776764000000003
        },
        "auroc": {
          "value": 0.5356010052080641,
          "lo": 0.5200740038979366,
          "hi": 0.5516852145424147
        },
        "acc@80": {
          "value": 0.39225,
          "lo": 0.3755,
          "hi": 0.40675
        }
      }
    },
    {
      "env": "boolq",
      "slice": "all",
      "system": "mix7-450M-s2",
      "metrics": {
        "n": {
          "value": 8129.0,
          "lo": 7984.475,
          "hi": 8270.575
        },
        "ce": {
          "value": 0.7508329721009208,
          "lo": 0.7358536079983113,
          "hi": 0.7664551576724759
        },
        "kl": {
          "value": 0.7508329721009208,
          "lo": 0.7358536079983113,
          "hi": 0.7664551576724759
        },
        "jsd": {
          "value": 0.20942762939546353,
          "lo": 0.20584896383359036,
          "hi": 0.2130091640498956
        },
        "brier": {
          "value": 0.5348680335227326,
          "lo": 0.5232469861530633,
          "hi": 0.5472430128018364
        },
        "acc": {
          "value": 0.5651371632427112,
          "lo": 0.5530735884451423,
          "hi": 0.575905109150735
        },
        "conf": {
          "value": 0.7431906673693836,
          "lo": 0.740020551374251,
          "hi": 0.7465173845156629
        },
        "ece": {
          "value": 0.1780535041266723,
          "lo": 0.16736057480662753,
          "hi": 0.1895971520382431
        },
        "reliability": {
          "value": 0.03491287469138259,
          "lo": 0.031337234242882296,
          "hi": 0.039678006536083436
        },
        "resolution": {
          "value": 0.013110929383974334,
          "lo": 0.011345961017750624,
          "hi": 0.0159477605978508
        },
        "uncertainty": {
          "value": 0.2457571499646924,
          "lo": 0.24423841095816076,
          "hi": 0.2471831856925081
        },
        "auroc": {
          "value": 0.6219674022878374,
          "lo": 0.6097153466302697,
          "hi": 0.6349263420934658
        },
        "acc@80": {
          "value": 0.5918806704597878,
          "lo": 0.5782243718484255,
          "hi": 0.6048652525465069
        }
      }
    },
    {
      "env": "chaosnli",
      "slice": "all",
      "system": "mix7-450M-s2",
      "metrics": {
        "n": {
          "value": 2763.0,
          "lo": 2732.475,
          "hi": 2798.525
        },
        "ce": {
          "value": 0.9068941962366361,
          "lo": 0.8888516301850681,
          "hi": 0.9230321660093607
        },
        "kl": {
          "value": 0.37653412616737425,
          "lo": 0.3590756962948154,
          "hi": 0.3930393550067818
        },
        "jsd": {
          "value": 0.0937054334217359,
          "lo": 0.08994113864237777,
          "hi": 0.09745959474970158
        },
        "brier": {
          "value": 0.2526296104819719,
          "lo": 0.24055254013616717,
          "hi": 0.26465888636043466
        },
        "acc": {
          "value": 0.5798045602605864,
          "lo": 0.5618061180489594,
          "hi": 0.5990943704525836
        },
        "conf": {
          "value": 0.642501125010381,
          "lo": 0.6378141976322809,
          "hi": 0.647348896192297
        },
        "ece": {
          "value": 0.08573853579112858,
          "lo": 0.0736734226832399,
          "hi": 0.10743256827409897
        },
        "reliability": {
          "value": 0.01218772358814043,
          "lo": 0.009377742264503228,
          "hi": 0.01817624780187894
        },
        "resolution": {
          "value": 0.004357190061475463,
          "lo": 0.0034560381761563077,
          "hi": 0.008340115579019522
        },
        "uncertainty": {
          "value": 0.24363123216161445,
          "lo": 0.24018030445412428,
          "hi": 0.24617989033616938
        },
        "auroc": {
          "value": 0.5601164995091192,
          "lo": 0.5395851619829724,
          "hi": 0.5819451806207654
        },
        "acc@80": {
          "value": 0.602262443438914,
          "lo": 0.5819354162148589,
          "hi": 0.6237648972781059
        }
      }
    },
    {
      "env": "civilcomments",
      "slice": "all",
      "system": "mix7-450M-s2",
      "metrics": {
        "n": {
          "value": 10581.0,
          "lo": 10261.425,
          "hi": 10904.525
        },
        "ce": {
          "value": 0.3247535194409554,
          "lo": 0.316974842136389,
          "hi": 0.3318598707986615
        },
        "kl": {
          "value": 0.09327071067161433,
          "lo": 0.09064621848006946,
          "hi": 0.0961752930409639
        },
        "jsd": {
          "value": 0.02606528108395236,
          "lo": 0.025350973955439525,
          "hi": 0.02679507349136137
        },
        "brier": {
          "value": 0.044632186213394674,
          "lo": 0.04271824048677188,
          "hi": 0.04659245419611563
        },
        "acc": {
          "value": 0.9398922597108024,
          "lo": 0.934916151775364,
          "hi": 0.9446732982719718
        },
        "conf": {
          "value": 0.8583259791454062,
          "lo": 0.8553666134039802,
          "hi": 0.8618230064894931
        },
        "ece": {
          "value": 0.08156628056539615,
          "lo": 0.07692101049155381,
          "hi": 0.08594324081028143
        },
        "reliability": {
          "value": 0.00975207879263131,
          "lo": 0.008637521595815463,
          "hi": 0.010896552611447664
        },
        "resolution": {
          "value": 0.010355037855520537,
          "lo": 0.008795191276641622,
          "hi": 0.012097052840457853
        },
        "uncertainty": {
          "value": 0.056494799846524,
          "lo": 0.05226565402111658,
          "hi": 0.06084793886484167
        },
        "auroc": {
          "value": 0.8993091721449102,
          "lo": 0.8904418349475692,
          "hi": 0.9084257351286236
        },
        "acc@80": {
          "value": 0.9846426461901949,
          "lo": 0.9819117897378651,
          "hi": 0.9873751089781088
        }
      }
    },
    {
      "env": "folktables",
      "slice": "all",
      "system": "mix7-450M-s2",
      "metrics": {
        "n": {
          "value": 15003.0,
          "lo": 14855.8,
          "hi": 15157.0
        },
        "ce": {
          "value": 0.4384948531105458,
          "lo": 0.4294832468311023,
          "hi": 0.4474398319850143
        },
        "kl": {
          "value": 0.4384948531105458,
          "lo": 0.4294832468311023,
          "hi": 0.4474398319850143
        },
        "jsd": {
          "value": 0.12703740170169311,
          "lo": 0.12481733848097287,
          "hi": 0.12926476086816707
        },
        "brier": {
          "value": 0.28842804801350497,
          "lo": 0.2818311021471689,
          "hi": 0.2952518595623806
        },
        "acc": {
          "value": 0.7847097247217223,
          "lo": 0.7777182306995476,
          "hi": 0.7918898360105096
        },
        "conf": {
          "value": 0.7962835168970399,
          "lo": 0.7936710437870885,
          "hi": 0.7987139368304775
        },
        "ece": {
          "value": 0.014699667606768206,
          "lo": 0.01157972838630544,
          "hi": 0.02266268690188863
        },
        "reliability": {
          "value": 0.00039307823531009166,
          "lo": 0.00023782628922572584,
          "hi": 0.0009264038908527687
        },
        "resolution": {
          "value": 0.0249705180780406,
          "lo": 0.02304098520361409,
          "hi": 0.027294810654141508
        },
        "uncertainty": {
          "value": 0.16894037264888112,
          "lo": 0.16480031233278103,
          "hi": 0.1728725842837765
        },
        "auroc": {
          "value": 0.7679356316954442,
          "lo": 0.7603918536552557,
          "hi": 0.7756067883381488
        },
        "acc@80": {
          "value": 0.8422762872854525,
          "lo": 0.8344919684580364,
          "hi": 0.849076861428603
        }
      }
    },
    {
      "env": "mcq",
      "slice": "all",
      "system": "mix7-450M-s2",
      "metrics": {
        "n": {
          "value": 5064.0,
          "lo": 5064.0,
          "hi": 5064.0
        },
        "ce": {
          "value": 0.7446275684259486,
          "lo": 0.718017758419548,
          "hi": 0.7752686950616806
        },
        "kl": {
          "value": 0.7446275684259486,
          "lo": 0.718017758419548,
          "hi": 0.7752686950616806
        },
        "jsd": {
          "value": 0.17277025351490105,
          "lo": 0.1681006773300853,
          "hi": 0.17823748981647008
        },
        "brier": {
          "value": 0.3832624003293176,
          "lo": 0.37073998564183014,
          "hi": 0.3978390423209227
        },
        "acc": {
          "value": 0.7278830963665087,
          "lo": 0.7161285545023697,
          "hi": 0.7391488941548183
        },
        "conf": {
          "value": 0.79739615916735,
          "lo": 0.7916504391466119,
          "hi": 0.8023993313515567
        },
        "ece": {
          "value": 0.06951306280084134,
          "lo": 0.059637876585902756,
          "hi": 0.07954645595597148
        },
        "reliability": {
          "value": 0.006545407602718125,
          "lo": 0.005115935871504275,
          "hi": 0.009169499928122849
        },
        "resolution": {
          "value": 0.04375928005498686,
          "lo": 0.0397319930327455,
          "hi": 0.04924385825133258
        },
        "uncertainty": {
          "value": 0.1980692943904125,
          "lo": 0.1928077675266853,
          "hi": 0.20328843820425568
        },
        "auroc": {
          "value": 0.800425865098159,
          "lo": 0.788415624556946,
          "hi": 0.8130848853706965
        },
        "acc@80": {
          "value": 0.8047395704764255,
          "lo": 0.7931374969143421,
          "hi": 0.8157306837817824
        }
      }
    },
    {
      "env": "stsb",
      "slice": "all",
      "system": "mix7-450M-s2",
      "metrics": {
        "n": {
          "value": 3676.0,
          "lo": 3150.95,
          "hi": 4793.674999999996
        },
        "ce": {
          "value": 1.1153752379761714,
          "lo": 1.052273713920008,
          "hi": 1.174999009193296
        },
        "kl": {
          "value": 0.7130714151139158,
          "lo": 0.6529503982280231,
          "hi": 0.7707624273377698
        },
        "jsd": {
          "value": 0.17712156250081665,
          "lo": 0.16455952178985314,
          "hi": 0.18869218834487786
        },
        "brier": {
          "value": 0.33239855464142265,
          "lo": 0.30888337549417993,
          "hi": 0.35571402833468196
        },
        "acc": {
          "value": 0.5424374319912949,
          "lo": 0.5119142542368635,
          "hi": 0.572282029555273
        },
        "conf": {
          "value": 0.5774308672505183,
          "lo": 0.567075724870965,
          "hi": 0.591399122720293
        },
        "ece": {
          "value": 0.040822435904119915,
          "lo": 0.03244471963777056,
          "hi": 0.06495263875134669
        },
        "reliability": {
          "value": 0.002350762670456841,
          "lo": 0.0015834587515945622,
          "hi": 0.005964010156505599
        },
        "resolution": {
          "value": 0.013943037849111384,
          "lo": 0.011632704935385763,
          "hi": 0.017970457315596322
        },
        "uncertainty": {
          "value": 0.2481990643659842,
          "lo": 0.2447753076204106,
          "hi": 0.24985804768021877
        },
        "auroc": {
          "value": 0.6181128104885405,
          "lo": 0.5952025262220156,
          "hi": 0.6359539621774973
        },
        "acc@80": {
          "value": 0.5702142128527712,
          "lo": 0.536247408370524,
          "hi": 0.6019217497658009
        }
      }
    },
    {
      "env": "yelp",
      "slice": "all",
      "system": "mix7-450M-s2",
      "metrics": {
        "n": {
          "value": 5000.0,
          "lo": 5000.0,
          "hi": 5000.0
        },
        "ce": {
          "value": 1.2990671352362657,
          "lo": 1.2826223818640479,
          "hi": 1.316544163088931
        },
        "kl": {
          "value": 1.2990671352362657,
          "lo": 1.2826223818640479,
          "hi": 1.316544163088931
        },
        "jsd": {
          "value": 0.3433666298098565,
          "lo": 0.34000996210799006,
          "hi": 0.34651044451864993
        },
        "brier": {
          "value": 0.6990184567747504,
          "lo": 0.6901114858907645,
          "hi": 0.7073333773478532
        },
        "acc": {
          "value": 0.3984,
          "lo": 0.384085,
          "hi": 0.4118
        },
        "conf": {
          "value": 0.4726655150473118,
          "lo": 0.4689257402203232,
          "hi": 0.4758530365492403
        },
        "ece": {
          "value": 0.07504558734297752,
          "lo": 0.06929116827212274,
          "hi": 0.09375943623945117
        },
        "reliability": {
          "value": 0.012138016938554872,
          "lo": 0.009904491197695403,
          "hi": 0.01594855405546701
        },
        "resolution": {
          "value": 0.005249490304021365,
          "lo": 0.0038488391332653954,
          "hi": 0.008041268008154764
        },
        "uncertainty": {
          "value": 0.23967744,
          "lo": 0.236563623,
          "hi": 0.24222076000000003
        },
        "auroc": {
          "value": 0.5681968231970435,
          "lo": 0.5512457926399172,
          "hi": 0.5857333428553422
        },
        "acc@80": {
          "value": 0.42575,
          "lo": 0.4105,
          "hi": 0.44125
        }
      }
    },
    {
      "env": "boolq",
      "slice": "all",
      "system": "mix7-3B-s0",
      "metrics": {
        "n": {
          "value": 8129.0,
          "lo": 7984.475,
          "hi": 8270.575
        },
        "ce": {
          "value": 0.4015344595969336,
          "lo": 0.3860123018209743,
          "hi": 0.4162750405789503
        },
        "kl": {
          "value": 0.4015344595969336,
          "lo": 0.3860123018209743,
          "hi": 0.4162750405789503
        },
        "jsd": {
          "value": 0.11153371062325852,
          "lo": 0.10802042083492568,
          "hi": 0.11504278763351378
        },
        "brier": {
          "value": 0.2557792622318184,
          "lo": 0.24457311973185,
          "hi": 0.26649771244224674
        },
        "acc": {
          "value": 0.8148603764300651,
          "lo": 0.8049397055807423,
          "hi": 0.8241163617325865
        },
        "conf": {
          "value": 0.8385382051000971,
          "lo": 0.8348062592992268,
          "hi": 0.842005228281196
        },
        "ece": {
          "value": 0.023677828670031976,
          "lo": 0.018086918447659944,
          "hi": 0.033850399052529526
        },
        "reliability": {
          "value": 0.0009166159624454497,
          "lo": 0.0005855294696240854,
          "hi": 0.0018833355470140156
        },
        "resolution": {
          "value": 0.02349288559872494,
          "lo": 0.021108489281470456,
          "hi": 0.02688732478426397
        },
        "uncertainty": {
          "value": 0.15086294335431769,
          "lo": 0.14494858385545084,
          "hi": 0.15701177536230493
        },
        "auroc": {
          "value": 0.7821920590784343,
          "lo": 0.7696202531580901,
          "hi": 0.794582540631279
        },
        "acc@80": {
          "value": 0.8782100568968169,
          "lo": 0.8675312644742936,
          "hi": 0.8860653702855604
        }
      }
    },
    {
      "env": "chaosnli",
      "slice": "all",
      "system": "mix7-3B-s0",
      "metrics": {
        "n": {
          "value": 2763.0,
          "lo": 2732.475,
          "hi": 2798.525
        },
        "ce": {
          "value": 0.7317921943962872,
          "lo": 0.7175234800133125,
          "hi": 0.7461812755771366
        },
        "kl": {
          "value": 0.2014321243270253,
          "lo": 0.1915986327201315,
          "hi": 0.2125726256760473
        },
        "jsd": {
          "value": 0.050709724346277796,
          "lo": 0.04841450046627548,
          "hi": 0.05330863049440536
        },
        "brier": {
          "value": 0.13177767445592195,
          "lo": 0.1241462323637343,
          "hi": 0.1397989300330173
        },
        "acc": {
          "value": 0.7321751719145856,
          "lo": 0.7150233697658727,
          "hi": 0.7488255251608713
        },
        "conf": {
          "value": 0.6842451099248369,
          "lo": 0.6790877096119059,
          "hi": 0.689623098340747
        },
        "ece": {
          "value": 0.048137297006435174,
          "lo": 0.03969545172528373,
          "hi": 0.06678392948039072
        },
        "reliability": {
          "value": 0.003264699173978197,
          "lo": 0.002404396547416299,
          "hi": 0.006193114064328454
        },
        "resolution": {
          "value": 0.021631081394147284,
          "lo": 0.018310029205294103,
          "hi": 0.02754729414658824
        },
        "uncertainty": {
          "value": 0.19609468954643264,
          "lo": 0.18808585802554073,
          "hi": 0.20376483602107143
        },
        "auroc": {
          "value": 0.7111210939065611,
          "lo": 0.6915367421288887,
          "hi": 0.7322106870758519
        },
        "acc@80": {
          "value": 0.783710407239819,
          "lo": 0.7668429473017556,
          "hi": 0.8015500885942087
        }
      }
    },
    {
      "env": "civilcomments",
      "slice": "all",
      "system": "mix7-3B-s0",
      "metrics": {
        "n": {
          "value": 10581.0,
          "lo": 10261.425,
          "hi": 10904.525
        },
        "ce": {
          "value": 0.32493529455042947,
          "lo": 0.31631111338723766,
          "hi": 0.332856731464495
        },
        "kl": {
          "value": 0.09345248578108843,
          "lo": 0.09049580801852516,
          "hi": 0.09640686219627244
        },
        "jsd": {
          "value": 0.023236853210795962,
          "lo": 0.022569521394711495,
          "hi": 0.023902308587587625
        },
        "brier": {
          "value": 0.044440062724136684,
          "lo": 0.0424329646687974,
          "hi": 0.04654394576966348
        },
        "acc": {
          "value": 0.9376240430961157,
          "lo": 0.9325575502601535,
          "hi": 0.9421614171606318
        },
        "conf": {
          "value": 0.8938100220361445,
          "lo": 0.8909953417291749,
          "hi": 0.8969365017590876
        },
        "ece": {
          "value": 0.045806885644400594,
          "lo": 0.04272027448871869,
          "hi": 0.049717523521862775
        },
        "reliability": {
          "value": 0.003369591571482668,
          "lo": 0.002855248382434492,
          "hi": 0.003956748929645738
        },
        "resolution": {
          "value": 0.013996847595005229,
          "lo": 0.011918922910384185,
          "hi": 0.016271498684583937
        },
        "uncertainty": {
          "value": 0.05848519690420908,
          "lo": 0.05449328087635878,
          "hi": 0.06289396568079476
        },
        "auroc": {
          "value": 0.9210726863433244,
          "lo": 0.9125079832657689,
          "hi": 0.9294178166364858
        },
        "acc@80": {
          "value": 0.9868871825162433,
          "lo": 0.9844468765677032,
          "hi": 0.9895334115529497
        }
      }
    },
    {
      "env": "folktables",
      "slice": "all",
      "system": "mix7-3B-s0",
      "metrics": {
        "n": {
          "value": 15003.0,
          "lo": 14855.8,
          "hi": 15157.0
        },
        "ce": {
          "value": 0.44509115726019133,
          "lo": 0.43615898719736423,
          "hi": 0.4543582098797216
        },
        "kl": {
          "value": 0.44509115726019133,
          "lo": 0.43615898719736423,
          "hi": 0.4543582098797216
        },
        "jsd": {
          "value": 0.12672637436350415,
          "lo": 0.12465214829656616,
          "hi": 0.12904600548269632
        },
        "brier": {
          "value": 0.2930003292994701,
          "lo": 0.28668903733955103,
          "hi": 0.3004513879616006
        },
        "acc": {
          "value": 0.7876424715056989,
          "lo": 0.7805317861129476,
          "hi": 0.7939606685270852
        },
        "conf": {
          "value": 0.8070413636770326,
          "lo": 0.8041954014356815,
          "hi": 0.8095585776890852
        },
        "ece": {
          "value": 0.02779868638635452,
          "lo": 0.022337478545157424,
          "hi": 0.0340497470428343
        },
        "reliability": {
          "value": 0.001027915756892152,
          "lo": 0.0007314647635620185,
          "hi": 0.0016595918413771723
        },
        "resolution": {
          "value": 0.02169735023637832,
          "lo": 0.020282061718895836,
          "hi": 0.02374580577398444
        },
        "uncertainty": {
          "value": 0.1672618085860932,
          "lo": 0.16358712397960315,
          "hi": 0.17130191668681444
        },
        "auroc": {
          "value": 0.7532512981367189,
          "lo": 0.7450062741277153,
          "hi": 0.761987673941276
        },
        "acc@80": {
          "value": 0.8386935510748209,
          "lo": 0.8307700002943383,
          "hi": 0.8448743534482759
        }
      }
    },
    {
      "env": "mcq",
      "slice": "all",
      "system": "mix7-3B-s0",
      "metrics": {
        "n": {
          "value": 5064.0,
          "lo": 5064.0,
          "hi": 5064.0
        },
        "ce": {
          "value": 0.28394581130659097,
          "lo": 0.26603788540962414,
          "hi": 0.3050438775655096
        },
        "kl": {
          "value": 0.28394581130659097,
          "lo": 0.26603788540962414,
          "hi": 0.3050438775655096
        },
        "jsd": {
          "value": 0.06816561102544832,
          "lo": 0.06459715585964569,
          "hi": 0.0721479748574227
        },
        "brier": {
          "value": 0.14707832873591406,
          "lo": 0.13726122315282865,
          "hi": 0.1568657334447049
        },
        "acc": {
          "value": 0.8992890995260664,
          "lo": 0.8912766587677725,
          "hi": 0.9069905213270142
        },
        "conf": {
          "value": 0.9096089693560231,
          "lo": 0.9052909305017577,
          "hi": 0.9133925156906176
        },
        "ece": {
          "value": 0.01118656555058266,
          "lo": 0.009677040375810336,
          "hi": 0.02007273871548722
        },
        "reliability": {
          "value": 0.00036228744679989394,
          "lo": 0.00023085968401454236,
          "hi": 0.001134681554697722
        },
        "resolution": {
          "value": 0.02394559257678946,
          "lo": 0.020469756669410887,
          "hi": 0.028077017761401986
        },
        "uncertainty": {
          "value": 0.09056821499966304,
          "lo": 0.08435871554996521,
          "hi": 0.09690248878338312
        },
        "auroc": {
          "value": 0.9004529523711109,
          "lo": 0.8896641740533964,
          "hi": 0.9115282046258159
        },
        "acc@80": {
          "value": 0.9674154529745742,
          "lo": 0.9612441372500617,
          "hi": 0.9728462108121452
        }
      }
    },
    {
      "env": "stsb",
      "slice": "all",
      "system": "mix7-3B-s0",
      "metrics": {
        "n": {
          "value": 3676.0,
          "lo": 3150.95,
          "hi": 4793.674999999996
        },
        "ce": {
          "value": 1.0320934313543655,
          "lo": 0.9737943844535284,
          "hi": 1.0907567401781189
        },
        "kl": {
          "value": 0.6297896084921099,
          "lo": 0.5771812394309076,
          "hi": 0.6852677359838089
        },
        "jsd": {
          "value": 0.14011290764967546,
          "lo": 0.13022503773077626,
          "hi": 0.14963892944676882
        },
        "brier": {
          "value": 0.2879302523060823,
          "lo": 0.26417715261978636,
          "hi": 0.31065636616289766
        },
        "acc": {
          "value": 0.6235038084874864,
          "lo": 0.5980606588589551,
          "hi": 0.6456225171251262
        },
        "conf": {
          "value": 0.6865655293564542,
          "lo": 0.6821742615009048,
          "hi": 0.6905385006638595
        },
        "ece": {
          "value": 0.06937201697302331,
          "lo": 0.05758364124893235,
          "hi": 0.09149823202160497
        },
        "reliability": {
          "value": 0.006808533333740608,
          "lo": 0.004744926620305128,
          "hi": 0.011386983417516254
        },
        "resolution": {
          "value": 0.012280547113417957,
          "lo": 0.009840436482986279,
          "hi": 0.01631852055148593
        },
        "uncertainty": {
          "value": 0.23474680928908628,
          "lo": 0.22879368686549964,
          "hi": 0.24038410235088176
        },
        "auroc": {
          "value": 0.6192176040815504,
          "lo": 0.5996142966251977,
          "hi": 0.6389117903837324
        },
        "acc@80": {
          "value": 0.6501190071404285,
          "lo": 0.622900699736252,
          "hi": 0.6762869012258145
        }
      }
    },
    {
      "env": "yelp",
      "slice": "all",
      "system": "mix7-3B-s0",
      "metrics": {
        "n": {
          "value": 5000.0,
          "lo": 5000.0,
          "hi": 5000.0
        },
        "ce": {
          "value": 1.0391471499843858,
          "lo": 1.0134554300556946,
          "hi": 1.0625187943526921
        },
        "kl": {
          "value": 1.0391471499843858,
          "lo": 1.0134554300556946,
          "hi": 1.0625187943526921
        },
        "jsd": {
          "value": 0.2630772132910125,
          "lo": 0.2576900758880535,
          "hi": 0.2682183793453246
        },
        "brier": {
          "value": 0.5788665208432717,
          "lo": 0.563677305243213,
          "hi": 0.5927627160220487
        },
        "acc": {
          "value": 0.5574,
          "lo": 0.5436949999999999,
          "hi": 0.571705
        },
        "conf": {
          "value": 0.6398223581820727,
          "lo": 0.6342675311058015,
          "hi": 0.6450378332845866
        },
        "ece": {
          "value": 0.08723655652105806,
          "lo": 0.07699143006928265,
          "hi": 0.10208049490526318
        },
        "reliability": {
          "value": 0.011098361250199508,
          "lo": 0.009022912139590654,
          "hi": 0.014836317323023748
        },
        "resolution": {
          "value": 0.019314236421571273,
          "lo": 0.01676013430935074,
          "hi": 0.023318424272665134
        },
        "uncertainty": {
          "value": 0.24670524,
          "lo": 0.24485838299999999,
          "hi": 0.248090737
        },
        "auroc": {
          "value": 0.650048049242894,
          "lo": 0.633433680844322,
          "hi": 0.6650909921020128
        },
        "acc@80": {
          "value": 0.59225,
          "lo": 0.57625,
          "hi": 0.6096312500000001
        }
      }
    },
    {
      "env": "boolq",
      "slice": "all",
      "system": "mix7-3B-s1",
      "metrics": {
        "n": {
          "value": 8129.0,
          "lo": 7984.475,
          "hi": 8270.575
        },
        "ce": {
          "value": 0.4049574630856653,
          "lo": 0.39007587939008853,
          "hi": 0.419696224864637
        },
        "kl": {
          "value": 0.4049574630856653,
          "lo": 0.39007587939008853,
          "hi": 0.419696224864637
        },
        "jsd": {
          "value": 0.111901545028253,
          "lo": 0.10854569451679102,
          "hi": 0.11533485690992372
        },
        "brier": {
          "value": 0.25681659383982436,
          "lo": 0.24637147985817529,
          "hi": 0.2673096727568516
        },
        "acc": {
          "value": 0.8173207036535859,
          "lo": 0.8080316120551718,
          "hi": 0.8275816552524383
        },
        "conf": {
          "value": 0.8375277741176204,
          "lo": 0.8336681292483965,
          "hi": 0.841110267633026
        },
        "ece": {
          "value": 0.020491637790873277,
          "lo": 0.016299947854084208,
          "hi": 0.03069316298435388
        },
        "reliability": {
          "value": 0.0007266617534696402,
          "lo": 0.00047220778666538565,
          "hi": 0.001557464036775021
        },
        "resolution": {
          "value": 0.02125015517324278,
          "lo": 0.018781052645828994,
          "hi": 0.023955819660760647
        },
        "uncertainty": {
          "value": 0.14930757103279313,
          "lo": 0.14269003212759473,
          "hi": 0.1551165258621473
        },
        "auroc": {
          "value": 0.7717524431552125,
          "lo": 0.759722393146674,
          "hi": 0.7835680600374859
        },
        "acc@80": {
          "value": 0.8769798554513302,
          "lo": 0.8675635293138366,
          "hi": 0.8852583732057416
        }
      }
    },
    {
      "env": "chaosnli",
      "slice": "all",
      "system": "mix7-3B-s1",
      "metrics": {
        "n": {
          "value": 2763.0,
          "lo": 2732.475,
          "hi": 2798.525
        },
        "ce": {
          "value": 0.7349974377531698,
          "lo": 0.7208577981229234,
          "hi": 0.7501208449342154
        },
        "kl": {
          "value": 0.204637367683908,
          "lo": 0.19473334216220856,
          "hi": 0.21571769509786942
        },
        "jsd": {
          "value": 0.05172847852322033,
          "lo": 0.04934382396665588,
          "hi": 0.05433960084166812
        },
        "brier": {
          "value": 0.13334485416438893,
          "lo": 0.12546612960853043,
          "hi": 0.14151518877841704
        },
        "acc": {
          "value": 0.7285559174809989,
          "lo": 0.7118080916048567,
          "hi": 0.7442746559724079
        },
        "conf": {
          "value": 0.687175090378538,
          "lo": 0.6818579669141838,
          "hi": 0.6933242753211061
        },
        "ece": {
          "value": 0.04855649530822884,
          "lo": 0.04099712850132304,
          "hi": 0.06923226641689625
        },
        "reliability": {
          "value": 0.003739122579211648,
          "lo": 0.0027199493628244638,
          "hi": 0.006664422066402824
        },
        "resolution": {
          "value": 0.023876320718372028,
          "lo": 0.019966860404635328,
          "hi": 0.030223145045202605
        },
        "uncertainty": {
          "value": 0.19776219258441882,
          "lo": 0.19032988663427408,
          "hi": 0.20513721000332208
        },
        "auroc": {
          "value": 0.7173164431197218,
          "lo": 0.6983102158074311,
          "hi": 0.7388496103712955
        },
        "acc@80": {
          "value": 0.783710407239819,
          "lo": 0.7663283716734326,
          "hi": 0.8005933503074701
        }
      }
    },
    {
      "env": "civilcomments",
      "slice": "all",
      "system": "mix7-3B-s1",
      "metrics": {
        "n": {
          "value": 10581.0,
          "lo": 10261.425,
          "hi": 10904.525
        },
        "ce": {
          "value": 0.31942936538814876,
          "lo": 0.31149573393331353,
          "hi": 0.3271525389195797
        },
        "kl": {
          "value": 0.0879465566188077,
          "lo": 0.08520028414159837,
          "hi": 0.09096505372521008
        },
        "jsd": {
          "value": 0.022278551457989914,
          "lo": 0.021592106042922107,
          "hi": 0.022963349235797917
        },
        "brier": {
          "value": 0.040234408557455296,
          "lo": 0.03843925900728954,
          "hi": 0.04211302940769042
        },
        "acc": {
          "value": 0.9445232019657878,
          "lo": 0.9398753114478805,
          "hi": 0.9490178170496311
        },
        "conf": {
          "value": 0.877979220542731,
          "lo": 0.8746057752849715,
          "hi": 0.8814415317216445
        },
        "ece": {
          "value": 0.06654398142305668,
          "lo": 0.062062010685223434,
          "hi": 0.07060110024882085
        },
        "reliability": {
          "value": 0.00795537710407801,
          "lo": 0.007002408528562913,
          "hi": 0.0090299316599831
        },
        "resolution": {
          "value": 0.010679213571236502,
          "lo": 0.009024108954011732,
          "hi": 0.012601257559940677
        },
        "uncertainty": {
          "value": 0.05239912291408346,
          "lo": 0.04838299931780303,
          "hi": 0.056509704103497885
        },
        "auroc": {
          "value": 0.913231925526696,
          "lo": 0.9052468216388458,
          "hi": 0.9223523500697764
        },
        "acc@80": {
          "value": 0.9890135853514471,
          "lo": 0.9868124670718026,
          "hi": 0.9914803908055612
        }
      }
    },
    {
      "env": "folktables",
      "slice": "all",
      "system": "mix7-3B-s1",
      "metrics": {
        "n": {
          "value": 15003.0,
          "lo": 14855.8,
          "hi": 15157.0
        },
        "ce": {
          "value": 0.44111304791622236,
          "lo": 0.43174813491236325,
          "hi": 0.44934647479135614
        },
        "kl": {
          "value": 0.44111304791622236,
          "lo": 0.43174813491236325,
          "hi": 0.44934647479135614
        },
        "jsd": {
          "value": 0.12770348267111706,
          "lo": 0.12550513145803793,
          "hi": 0.12981442786045344
        },
        "brier": {
          "value": 0.289130483213222,
          "lo": 0.2827122997681443,
          "hi": 0.2951485319528612
        },
        "acc": {
          "value": 0.7869759381457042,
          "lo": 0.779830205792653,
          "hi": 0.7937276150183592
        },
        "conf": {
          "value": 0.7849255870844709,
          "lo": 0.7818482000939939,
          "hi": 0.7876585101401754
        },
        "ece": {
          "value": 0.014096905742098072,
          "lo": 0.011880030806329907,
          "hi": 0.021862798972842573
        },
        "reliability": {
          "value": 0.0002827487042510449,
          "lo": 0.0002090908089736323,
          "hi": 0.0006698803904785933
        },
        "resolution": {
          "value": 0.02329595916455912,
          "lo": 0.02145050436541092,
          "hi": 0.025442393969833454
        },
        "uncertainty": {
          "value": 0.16764481092539296,
          "lo": 0.16372408614990336,
          "hi": 0.17169505174081806
        },
        "auroc": {
          "value": 0.7613656299221321,
          "lo": 0.7528968817560899,
          "hi": 0.7702832216209622
        },
        "acc@80": {
          "value": 0.8421096483919347,
          "lo": 0.8348447237671797,
          "hi": 0.8485784365383892
        }
      }
    },
    {
      "env": "mcq",
      "slice": "all",
      "system": "mix7-3B-s1",
      "metrics": {
        "n": {
          "value": 5064.0,
          "lo": 5064.0,
          "hi": 5064.0
        },
        "ce": {
          "value": 0.2891957645267851,
          "lo": 0.27197159366670415,
          "hi": 0.30715282098789554
        },
        "kl": {
          "value": 0.2891957645267851,
          "lo": 0.27197159366670415,
          "hi": 0.30715282098789554
        },
        "jsd": {
          "value": 0.07248578135348922,
          "lo": 0.0688045261914869,
          "hi": 0.07633694894831154
        },
        "brier": {
          "value": 0.1517139060824241,
          "lo": 0.14185286472944145,
          "hi": 0.16164069421183636
        },
        "acc": {
          "value": 0.8937598736176935,
          "lo": 0.8855598341232228,
          "hi": 0.9019599131121643
        },
        "conf": {
          "value": 0.8991462239086345,
          "lo": 0.8944063981877284,
          "hi": 0.9031266109305794
        },
        "ece": {
          "value": 0.00769100147542233,
          "lo": 0.006277996731118698,
          "hi": 0.01696278155835766
        },
        "reliability": {
          "value": 0.00025283248472168286,
          "lo": 9.270644217381292e-05,
          "hi": 0.0009927991395790466
        },
        "resolution": {
          "value": 0.027283589988228798,
          "lo": 0.023354736970078813,
          "hi": 0.031692906294089315
        },
        "uncertainty": {
          "value": 0.09495316192857806,
          "lo": 0.08842821852640076,
          "hi": 0.1013436045864124
        },
        "auroc": {
          "value": 0.9045484002385228,
          "lo": 0.8941789210892567,
          "hi": 0.9148721676895425
        },
        "acc@80": {
          "value": 0.9696371266353987,
          "lo": 0.9637126635398667,
          "hi": 0.9743273265860282
        }
      }
    },
    {
      "env": "stsb",
      "slice": "all",
      "system": "mix7-3B-s1",
      "metrics": {
        "n": {
          "value": 3676.0,
          "lo": 3150.95,
          "hi": 4793.674999999996
        },
        "ce": {
          "value": 0.9606397550106914,
          "lo": 0.9136520337197678,
          "hi": 1.0053183174223232
        },
        "kl": {
          "value": 0.5583359321484358,
          "lo": 0.5134575243067591,
          "hi": 0.5998446529037125
        },
        "jsd": {
          "value": 0.1417157022674515,
          "lo": 0.13153584188974404,
          "hi": 0.15075761011863018
        },
        "brier": {
          "value": 0.2696511629637947,
          "lo": 0.25035290256773923,
          "hi": 0.2883582189865905
        },
        "acc": {
          "value": 0.6109902067464635,
          "lo": 0.5904998647433761,
          "hi": 0.62975556849448
        },
        "conf": {
          "value": 0.6308347471342123,
          "lo": 0.621701280045729,
          "hi": 0.6416677557116188
        },
        "ece": {
          "value": 0.03286231362683471,
          "lo": 0.027581300777750294,
          "hi": 0.053216983102494
        },
        "reliability": {
          "value": 0.0019507831759346804,
          "lo": 0.0013248013918855113,
          "hi": 0.004471052294358259
        },
        "resolution": {
          "value": 0.01910747870379169,
          "lo": 0.01575617488715062,
          "hi": 0.02309175560386782
        },
        "uncertainty": {
          "value": 0.23768117400637728,
          "lo": 0.23316348711925408,
          "hi": 0.24180977137399856
        },
        "auroc": {
          "value": 0.6517820336386676,
          "lo": 0.6311177764729875,
          "hi": 0.6689101056191983
        },
        "acc@80": {
          "value": 0.6501190071404285,
          "lo": 0.627964313123561,
          "hi": 0.6712623429020483
        }
      }
    },
    {
      "env": "yelp",
      "slice": "all",
      "system": "mix7-3B-s1",
      "metrics": {
        "n": {
          "value": 5000.0,
          "lo": 5000.0,
          "hi": 5000.0
        },
        "ce": {
          "value": 0.9725467655960507,
          "lo": 0.9492864944117893,
          "hi": 0.9923747008198048
        },
        "kl": {
          "value": 0.9725467655960507,
          "lo": 0.9492864944117893,
          "hi": 0.9923747008198048
        },
        "jsd": {
          "value": 0.2595922501934123,
          "lo": 0.2547513860074246,
          "hi": 0.2642575432437123
        },
        "brier": {
          "value": 0.5467747525784498,
          "lo": 0.5338918415026284,
          "hi": 0.5591425403359902
        },
        "acc": {
          "value": 0.5768,
          "lo": 0.5646,
          "hi": 0.59242
        },
        "conf": {
          "value": 0.6021929107874632,
          "lo": 0.5970300256822259,
          "hi": 0.6071611789501459
        },
        "ece": {
          "value": 0.035365436366200466,
          "lo": 0.030351453499123455,
          "hi": 0.0507533370282501
        },
        "reliability": {
          "value": 0.002083417082839355,
          "lo": 0.0016458693872491098,
          "hi": 0.003948515192838469
        },
        "resolution": {
          "value": 0.019561498181295066,
          "lo": 0.01691330508509431,
          "hi": 0.023520177786882694
        },
        "uncertainty": {
          "value": 0.24410176,
          "lo": 0.241458384,
          "hi": 0.24582684
        },
        "auroc": {
          "value": 0.6583241349837051,
          "lo": 0.6431085844867885,
          "hi": 0.6750408391933727
        },
        "acc@80": {
          "value": 0.6155,
          "lo": 0.59975,
          "hi": 0.632
        }
      }
    },
    {
      "env": "boolq",
      "slice": "all",
      "system": "mix7-3B-s2",
      "metrics": {
        "n": {
          "value": 8129.0,
          "lo": 7984.475,
          "hi": 8270.575
        },
        "ce": {
          "value": 0.4131032401681462,
          "lo": 0.3973496692894577,
          "hi": 0.427998296376232
        },
        "kl": {
          "value": 0.4131032401681462,
          "lo": 0.3973496692894577,
          "hi": 0.427998296376232
        },
        "jsd": {
          "value": 0.11348010140326135,
          "lo": 0.10986484531222152,
          "hi": 0.11709720283232482
        },
        "brier": {
          "value": 0.26307555605300936,
          "lo": 0.25154510471816366,
          "hi": 0.27411998163707557
        },
        "acc": {
          "value": 0.8106778201500799,
          "lo": 0.801569355404868,
          "hi": 0.8200104361128396
        },
        "conf": {
          "value": 0.8402179098830245,
          "lo": 0.8364491696032379,
          "hi": 0.8435558076027064
        },
        "ece": {
          "value": 0.02967430733860104,
          "lo": 0.022061073993597972,
          "hi": 0.03808999907093463
        },
        "reliability": {
          "value": 0.001150064358198883,
          "lo": 0.0007762966791525043,
          "hi": 0.002134593047143699
        },
        "resolution": {
          "value": 0.02284729456198637,
          "lo": 0.02037014493284156,
          "hi": 0.02564082887246402
        },
        "uncertainty": {
          "value": 0.1534792920667946,
          "lo": 0.14759312013249243,
          "hi": 0.15905592089447454
        },
        "auroc": {
          "value": 0.7737325244207016,
          "lo": 0.7613846273950495,
          "hi": 0.7844834670571107
        },
        "acc@80": {
          "value": 0.8726741503921267,
          "lo": 0.8627757865114729,
          "hi": 0.8818397680978345
        }
      }
    },
    {
      "env": "chaosnli",
      "slice": "all",
      "system": "mix7-3B-s2",
      "metrics": {
        "n": {
          "value": 2763.0,
          "lo": 2732.475,
          "hi": 2798.525
        },
        "ce": {
          "value": 0.7314480272773598,
          "lo": 0.7165841531560584,
          "hi": 0.7462843837267661
        },
        "kl": {
          "value": 0.2010879572080979,
          "lo": 0.19093257237017797,
          "hi": 0.21267601193959595
        },
        "jsd": {
          "value": 0.05125472097302959,
          "lo": 0.048962697249526135,
          "hi": 0.053971984200325965
        },
        "brier": {
          "value": 0.13283454092153876,
          "lo": 0.12531100763357664,
          "hi": 0.14145555307972574
        },
        "acc": {
          "value": 0.7350705754614549,
          "lo": 0.7182138184335632,
          "hi": 0.7539765682599965
        },
        "conf": {
          "value": 0.6815311866871395,
          "lo": 0.6761925889266218,
          "hi": 0.6868314750200839
        },
        "ece": {
          "value": 0.054777696259505855,
          "lo": 0.04402893164749189,
          "hi": 0.07315946449562966
        },
        "reliability": {
          "value": 0.00425957846132254,
          "lo": 0.003025259707686656,
          "hi": 0.007968014331708758
        },
        "resolution": {
          "value": 0.019088587588990097,
          "lo": 0.015589169144610369,
          "hi": 0.0244941432336972
        },
        "uncertainty": {
          "value": 0.19474182455222042,
          "lo": 0.1854955533308363,
          "hi": 0.20238271613029457
        },
        "auroc": {
          "value": 0.698795715588703,
          "lo": 0.6777626810697348,
          "hi": 0.7203468860875333
        },
        "acc@80": {
          "value": 0.783710407239819,
          "lo": 0.7646371632765255,
          "hi": 0.8012821438681834
        }
      }
    },
    {
      "env": "civilcomments",
      "slice": "all",
      "system": "mix7-3B-s2",
      "metrics": {
        "n": {
          "value": 10581.0,
          "lo": 10261.425,
          "hi": 10904.525
        },
        "ce": {
          "value": 0.3162963367925811,
          "lo": 0.30857301423871375,
          "hi": 0.32356174573071533
        },
        "kl": {
          "value": 0.08481352802324012,
          "lo": 0.08228301201192431,
          "hi": 0.087507702002961
        },
        "jsd": {
          "value": 0.02294056032259847,
          "lo": 0.02228275480633488,
          "hi": 0.02362822533140682
        },
        "brier": {
          "value": 0.0392586704145921,
          "lo": 0.037557024866026884,
          "hi": 0.040924470986403415
        },
        "acc": {
          "value": 0.9453737831962953,
          "lo": 0.9408058678375908,
          "hi": 0.9497721239851789
        },
        "conf": {
          "value": 0.8633439165355465,
          "lo": 0.8598479996963536,
          "hi": 0.8670561465722972
        },
        "ece": {
          "value": 0.08202986666074875,
          "lo": 0.07739240662636059,
          "hi": 0.0860125247996312
        },
        "reliability": {
          "value": 0.011687426094487264,
          "lo": 0.01055066906885218,
          "hi": 0.012810255865789797
        },
        "resolution": {
          "value": 0.010853550782185535,
          "lo": 0.009294702048276536,
          "hi": 0.012908751438985054
        },
        "uncertainty": {
          "value": 0.05164219324141937,
          "lo": 0.04770503161703058,
          "hi": 0.055690186487304075
        },
        "auroc": {
          "value": 0.9157116186943225,
          "lo": 0.9075945311556154,
          "hi": 0.9243755856561098
        },
        "acc@80": {
          "value": 0.9898405197873598,
          "lo": 0.9877442295024116,
          "hi": 0.9921699911982781
        }
      }
    },
    {
      "env": "folktables",
      "slice": "all",
      "system": "mix7-3B-s2",
      "metrics": {
        "n": {
          "value": 15003.0,
          "lo": 14855.8,
          "hi": 15157.0
        },
        "ce": {
          "value": 0.45390238176418163,
          "lo": 0.44257713882119504,
          "hi": 0.46506275189519375
        },
        "kl": {
          "value": 0.45390238176418163,
          "lo": 0.44257713882119504,
          "hi": 0.46506275189519375
        },
        "jsd": {
          "value": 0.1233984339381717,
          "lo": 0.12091792149914202,
          "hi": 0.12588460310110852
        },
        "brier": {
          "value": 0.2935044644711353,
          "lo": 0.2862414068257359,
          "hi": 0.30096003595232784
        },
        "acc": {
          "value": 0.7881090448576952,
          "lo": 0.7811839122725273,
          "hi": 0.7946900863642335
        },
        "conf": {
          "value": 0.8258664168088204,
          "lo": 0.8230876402888855,
          "hi": 0.8284361155496249
        },
        "ece": {
          "value": 0.03934656827201734,
          "lo": 0.035264071697279034,
          "hi": 0.046831063758928815
        },
        "reliability": {
          "value": 0.0023381925082443748,
          "lo": 0.0018911525496485661,
          "hi": 0.003290204609039463
        },
        "resolution": {
          "value": 0.022477558484952493,
          "lo": 0.02105558637455978,
          "hi": 0.024568491105061697
        },
        "uncertainty": {
          "value": 0.1669931782711866,
          "lo": 0.16315775020668882,
          "hi": 0.17093560734410718
        },
        "auroc": {
          "value": 0.7578438227483217,
          "lo": 0.7501047484842124,
          "hi": 0.7660957165442223
        },
        "acc@80": {
          "value": 0.8403599400099984,
          "lo": 0.832566749894723,
          "hi": 0.8466900172226277
        }
      }
    },
    {
      "env": "mcq",
      "slice": "all",
      "system": "mix7-3B-s2",
      "metrics": {
        "n": {
          "value": 5064.0,
          "lo": 5064.0,
          "hi": 5064.0
        },
        "ce": {
          "value": 0.2861016989829702,
          "lo": 0.2676710966396172,
          "hi": 0.3077958136689403
        },
        "kl": {
          "value": 0.2861016989829702,
          "lo": 0.2676710966396172,
          "hi": 0.3077958136689403
        },
        "jsd": {
          "value": 0.06600847510575883,
          "lo": 0.06225906479814482,
          "hi": 0.06988313589145256
        },
        "brier": {
          "value": 0.1476629415774208,
          "lo": 0.13761717537202953,
          "hi": 0.15737659063233683
        },
        "acc": {
          "value": 0.8996840442338072,
          "lo": 0.8914840047393365,
          "hi": 0.9069905213270142
        },
        "conf": {
          "value": 0.9201619169346718,
          "lo": 0.9160170786120998,
          "hi": 0.9236272624636431
        },
        "ece": {
          "value": 0.02074630266588068,
          "lo": 0.015070425714167513,
          "hi": 0.027729098887387272
        },
        "reliability": {
          "value": 0.0009633181351610409,
          "lo": 0.0005118152598797653,
          "hi": 0.0019473326818066304
        },
        "resolution": {
          "value": 0.0237194446043811,
          "lo": 0.020530420995619435,
          "hi": 0.027802868385090936
        },
        "uncertainty": {
          "value": 0.09025266478490804,
          "lo": 0.08435871554996521,
          "hi": 0.09674026430879061
        },
        "auroc": {
          "value": 0.9013879335375001,
          "lo": 0.8896879285924512,
          "hi": 0.9122964877806818
        },
        "acc@80": {
          "value": 0.9686497161194767,
          "lo": 0.9623488027647494,
          "hi": 0.9743273265860282
        }
      }
    },
    {
      "env": "stsb",
      "slice": "all",
      "system": "mix7-3B-s2",
      "metrics": {
        "n": {
          "value": 3676.0,
          "lo": 3150.95,
          "hi": 4793.674999999996
        },
        "ce": {
          "value": 1.0124180703390282,
          "lo": 0.9578298609416918,
          "hi": 1.0643383631981576
        },
        "kl": {
          "value": 0.6101142474767726,
          "lo": 0.559022520022428,
          "hi": 0.6599987950545193
        },
        "jsd": {
          "value": 0.13817419605804432,
          "lo": 0.12871165392582945,
          "hi": 0.1469010911344988
        },
        "brier": {
          "value": 0.28177827479458084,
          "lo": 0.2600940330569067,
          "hi": 0.30225653590456275
        },
        "acc": {
          "value": 0.6191512513601741,
          "lo": 0.5956881764551564,
          "hi": 0.6384247164223111
        },
        "conf": {
          "value": 0.6741650431245663,
          "lo": 0.6692797965312662,
          "hi": 0.679254919631878
        },
        "ece": {
          "value": 0.0637306943779023,
          "lo": 0.05506032962232484,
          "hi": 0.08412400401234231
        },
        "reliability": {
          "value": 0.00621485549369787,
          "lo": 0.004667391451410416,
          "hi": 0.010454610787192975
        },
        "resolution": {
          "value": 0.015622626577521078,
          "lo": 0.012215964784041453,
          "hi": 0.01979589089674316
        },
        "uncertainty": {
          "value": 0.2358029792993046,
          "lo": 0.23083859508637722,
          "hi": 0.24084376999801121
        },
        "auroc": {
          "value": 0.6282491212653779,
          "lo": 0.603501593465401,
          "hi": 0.6475211116319775
        },
        "acc@80": {
          "value": 0.6558993539612377,
          "lo": 0.6289360689362621,
          "hi": 0.6771410206860542
        }
      }
    },
    {
      "env": "yelp",
      "slice": "all",
      "system": "mix7-3B-s2",
      "metrics": {
        "n": {
          "value": 5000.0,
          "lo": 5000.0,
          "hi": 5000.0
        },
        "ce": {
          "value": 1.0011337405188199,
          "lo": 0.9753346307884541,
          "hi": 1.024778413367375
        },
        "kl": {
          "value": 1.0011337405188199,
          "lo": 0.9753346307884541,
          "hi": 1.024778413367375
        },
        "jsd": {
          "value": 0.2552398454410999,
          "lo": 0.24980107122494624,
          "hi": 0.26006344367899425
        },
        "brier": {
          "value": 0.5655031047936078,
          "lo": 0.5503120153958699,
          "hi": 0.5792040081646198
        },
        "acc": {
          "value": 0.5726,
          "lo": 0.55879,
          "hi": 0.58741
        },
        "conf": {
          "value": 0.649690758112073,
          "lo": 0.6439095934392512,
          "hi": 0.6554846212086826
        },
        "ece": {
          "value": 0.08159133845865724,
          "lo": 0.07506652643799783,
          "hi": 0.09863834241949022
        },
        "reliability": {
          "value": 0.011495647789278435,
          "lo": 0.009415492261957151,
          "hi": 0.01509313122028412
        },
        "resolution": {
          "value": 0.016553226261710813,
          "lo": 0.014624989329846983,
          "hi": 0.020412400144338892
        },
        "uncertainty": {
          "value": 0.24472924,
          "lo": 0.242359452,
          "hi": 0.246543696
        },
        "auroc": {
          "value": 0.6421032484716579,
          "lo": 0.6284965560499978,
          "hi": 0.6575292856478184
        },
        "acc@80": {
          "value": 0.60775,
          "lo": 0.59211875,
          "hi": 0.62563125
        }
      }
    }
  ],
  "paired_gate_results": {
    "450M": [
      {
        "tier": "B",
        "env": "mcq",
        "seed": 0,
        "ce_single": 0.7381567415998964,
        "ce": 0.747297009316037,
        "diff": {
          "value": 0.009140267716140581,
          "lo": -0.006196410089797683,
          "hi": 0.024727763838233493
        },
        "pass": true,
        "ce_base": 0.9293923372332724,
        "diff_base": {
          "value": -0.18209532791723532,
          "lo": -0.20456379334472086,
          "hi": -0.1581666467099493
        },
        "acc": 0.7156398104265402,
        "auroc": 0.8112850096578367,
        "ece": 0.06986760207721124,
        "conf": 0.7855074125037516,
        "flip": 0.16686413902053712,
        "brier": 0.38947030443786845
      },
      {
        "tier": "B",
        "env": "mcq",
        "seed": 1,
        "ce_single": 0.7256265333493006,
        "ce": 0.7268390369943714,
        "diff": {
          "value": 0.0012125036450708123,
          "lo": -0.016725134202668655,
          "hi": 0.019177147258237007
        },
        "pass": true,
        "ce_base": 0.9293923372332724,
        "diff_base": {
          "value": -0.20255330023890092,
          "lo": -0.22482715328826855,
          "hi": -0.17969118748438198
        },
        "acc": 0.7215639810426541,
        "auroc": 0.8156515544996836,
        "ece": 0.06707049818565308,
        "conf": 0.7886344792283071,
        "flip": 0.14533965244865718,
        "brier": 0.38048971460670034
      },
      {
        "tier": "B",
        "env": "mcq",
        "seed": 2,
        "ce_single": 0.7281025572024337,
        "ce": 0.7446275684259486,
        "diff": {
          "value": 0.016525011223514807,
          "lo": -0.000855911760031846,
          "hi": 0.034942436657513926
        },
        "pass": true,
        "ce_base": 0.9293923372332724,
        "diff_base": {
          "value": -0.18476476880732384,
          "lo": -0.20814323123273681,
          "hi": -0.1604615926601847
        },
        "acc": 0.7278830963665087,
        "auroc": 0.800425865098159,
        "ece": 0.06951306280084134,
        "conf": 0.79739615916735,
        "flip": 0.15916271721958924,
        "brier": 0.3832624003293176
      },
      {
        "tier": "B",
        "env": "chaosnli",
        "seed": 0,
        "ce_single": 0.9844640133367495,
        "ce": 0.8982318895558363,
        "diff": {
          "value": -0.08623212378091323,
          "lo": -0.10652471262333044,
          "hi": -0.06621382766939787
        },
        "pass": true,
        "ce_base": 0.9411396746035926,
        "diff_base": {
          "value": -0.042907785047756254,
          "lo": -0.05749265348496446,
          "hi": -0.028103012772523146
        },
        "acc": 0.5696706478465436,
        "auroc": 0.5786139997841288,
        "ece": 0.07636093434284788,
        "conf": 0.6321799852064542,
        "flip": 0.1237785016286645,
        "brier": 0.24486443443340125
      },
      {
        "tier": "B",
        "env": "chaosnli",
        "seed": 1,
        "ce_single": 0.8871500921895148,
        "ce": 0.9017071316185276,
        "diff": {
          "value": 0.014557039429012689,
          "lo": -0.0004969397159415857,
          "hi": 0.029173974675597524
        },
        "pass": true,
        "ce_base": 0.9411396746035926,
        "diff_base": {
          "value": -0.03943254298506512,
          "lo": -0.055159777065724384,
          "hi": -0.023642907355823452
        },
        "acc": 0.5754614549402823,
        "auroc": 0.5909461843255213,
        "ece": 0.07894042577598563,
        "conf": 0.6433491620774119,
        "flip": 0.1610568222946073,
        "brier": 0.24518685638804466
      },
      {
        "tier": "B",
        "env": "chaosnli",
        "seed": 2,
        "ce_single": 0.8954622802568517,
        "ce": 0.9068941962366361,
        "diff": {
          "value": 0.011431915979784435,
          "lo": -0.006660753236571492,
          "hi": 0.028242286674602532
        },
        "pass": true,
        "ce_base": 0.9411396746035926,
        "diff_base": {
          "value": -0.034245478366956496,
          "lo": -0.04885892896992744,
          "hi": -0.020397057125051746
        },
        "acc": 0.5798045602605864,
        "auroc": 0.5601164995091192,
        "ece": 0.08573853579112858,
        "conf": 0.642501125010381,
        "flip": 0.1986970684039088,
        "brier": 0.2526296104819719
      },
      {
        "tier": "B",
        "env": "cifar10",
        "seed": 0,
        "ce_single": 0.5567829186931413,
        "ce": 0.5611000733190353,
        "diff": {
          "value": 0.0043171546258939705,
          "lo": -0.02197660365861151,
          "hi": 0.03097862715820516
        },
        "pass": true,
        "ce_base": 0.8037275317717176,
        "diff_base": {
          "value": -0.24262745845268224,
          "lo": -0.30102144441524353,
          "hi": -0.18111650743956448
        },
        "acc": 0.8472521739130435,
        "auroc": 0.9243227687215418,
        "ece": 0.05521569651339365,
        "conf": 0.9024678704264372,
        "flip": 0.06149565217391304,
        "brier": 0.2198787598683563
      },
      {
        "tier": "B",
        "env": "cifar10",
        "seed": 1,
        "ce_single": 0.5469380003104549,
        "ce": 0.5091627016145139,
        "diff": {
          "value": -0.037775298695941054,
          "lo": -0.05927893644139918,
          "hi": -0.017854071484835844
        },
        "pass": true,
        "ce_base": 0.8037275317717176,
        "diff_base": {
          "value": -0.29456483015720364,
          "lo": -0.36004579887557103,
          "hi": -0.23134719299077566
        },
        "acc": 0.8509217391304348,
        "auroc": 0.9181642373285862,
        "ece": 0.04019794063930929,
        "conf": 0.8910135569981906,
        "flip": 0.08700869565217391,
        "brier": 0.21224998635895223
      },
      {
        "tier": "B",
        "env": "cifar10",
        "seed": 2,
        "ce_single": 0.5655811216366115,
        "ce": 0.58316263146589,
        "diff": {
          "value": 0.017581509829278405,
          "lo": -0.004978112560093911,
          "hi": 0.04073811228309107
        },
        "pass": true,
        "ce_base": 0.8037275317717176,
        "diff_base": {
          "value": -0.2205649003058277,
          "lo": -0.2781528600006001,
          "hi": -0.16262958098419275
        },
        "acc": 0.8423652173913043,
        "auroc": 0.9255523829062392,
        "ece": 0.0568483020759147,
        "conf": 0.8992134633206803,
        "flip": 0.05718260869565217,
        "brier": 0.22618299444310025
      },
      {
        "tier": "B",
        "env": "civilcomments",
        "seed": 0,
        "ce_single": 0.3190687141729324,
        "ce": 0.3362536541996374,
        "diff": {
          "value": 0.01718494002670492,
          "lo": 0.015083710602497,
          "hi": 0.01932160736486855
        },
        "pass": false,
        "ce_base": 0.5814737265683071,
        "diff_base": {
          "value": -0.2452200723686696,
          "lo": -0.25360402874176485,
          "hi": -0.23702421070301385
        },
        "acc": 0.9291182307910405,
        "auroc": 0.9130753060048147,
        "ece": 0.03717635887555369,
        "conf": 0.8974095615014813,
        "flip": null,
        "brier": 0.05199782543309763
      },
      {
        "tier": "B",
        "env": "civilcomments",
        "seed": 1,
        "ce_single": 0.3173171481792371,
        "ce": 0.32425900287305687,
        "diff": {
          "value": 0.006941854693819749,
          "lo": 0.005367884280422213,
          "hi": 0.008546991347228855
        },
        "pass": false,
        "ce_base": 0.5814737265683071,
        "diff_base": {
          "value": -0.2572147236952501,
          "lo": -0.2651383155764999,
          "hi": -0.24950532969871045
        },
        "acc": 0.9396087326339665,
        "auroc": 0.9061359484383446,
        "ece": 0.07329864480434391,
        "conf": 0.8663100878296226,
        "flip": null,
        "brier": 0.044416175920525686
      },
      {
        "tier": "B",
        "env": "civilcomments",
        "seed": 2,
        "ce_single": 0.3207698949153425,
        "ce": 0.3247535194409554,
        "diff": {
          "value": 0.00398362452561285,
          "lo": 0.0023974518068756983,
          "hi": 0.005536747774107673
        },
        "pass": false,
        "ce_base": 0.5814737265683071,
        "diff_base": {
          "value": -0.2567202071273516,
          "lo": -0.26446706818505655,
          "hi": -0.24923490914482513
        },
        "acc": 0.9398922597108024,
        "auroc": 0.8993091721449102,
        "ece": 0.08156628056539615,
        "conf": 0.8583259791454062,
        "flip": null,
        "brier": 0.044632186213394674
      },
      {
        "tier": "B",
        "env": "stsb",
        "seed": 0,
        "ce_single": 1.1361109137841074,
        "ce": 1.1119125656921782,
        "diff": {
          "value": -0.02419834809192916,
          "lo": -0.04609115024581878,
          "hi": 0.0016605323503595525
        },
        "pass": true,
        "ce_base": 1.7818855315091495,
        "diff_base": {
          "value": -0.669972965816971,
          "lo": -0.7339895191263983,
          "hi": -0.6063226809511448
        },
        "acc": 0.5525027203482046,
        "auroc": 0.6062123708655655,
        "ece": 0.04876845997228197,
        "conf": 0.5877989736487997,
        "flip": null,
        "brier": 0.3273993070991984
      },
      {
        "tier": "B",
        "env": "stsb",
        "seed": 1,
        "ce_single": 1.1748568926745255,
        "ce": 1.1069691649903395,
        "diff": {
          "value": -0.0678877276841858,
          "lo": -0.09197059477343385,
          "hi": -0.03933164131398174
        },
        "pass": true,
        "ce_base": 1.7818855315091495,
        "diff_base": {
          "value": -0.6749163665188097,
          "lo": -0.7423604856159355,
          "hi": -0.611100239422524
        },
        "acc": 0.5620239390642002,
        "auroc": 0.6223150024351675,
        "ece": 0.04551832948662383,
        "conf": 0.6037559004194459,
        "flip": null,
        "brier": 0.32481970890073175
      },
      {
        "tier": "B",
        "env": "stsb",
        "seed": 2,
        "ce_single": 1.226978893861858,
        "ce": 1.1153752379761714,
        "diff": {
          "value": -0.11160365588568648,
          "lo": -0.1390004700182866,
          "hi": -0.08355764749961739
        },
        "pass": true,
        "ce_base": 1.7818855315091495,
        "diff_base": {
          "value": -0.6665102935329779,
          "lo": -0.7348370252655096,
          "hi": -0.6021282107648279
        },
        "acc": 0.5424374319912949,
        "auroc": 0.6181128104885405,
        "ece": 0.040822435904119915,
        "conf": 0.5774308672505183,
        "flip": null,
        "brier": 0.33239855464142265
      },
      {
        "tier": "B",
        "env": "folktables",
        "seed": 0,
        "ce_single": 0.44187247322055384,
        "ce": 0.43932493729908706,
        "diff": {
          "value": -0.0025475359214667897,
          "lo": -0.00612057476814358,
          "hi": 0.0010668040060764063
        },
        "pass": true,
        "ce_base": 0.6934844593415295,
        "diff_base": {
          "value": -0.25415952204244235,
          "lo": -0.26311984755805745,
          "hi": -0.24539613398456228
        },
        "acc": 0.7843098047057255,
        "auroc": 0.7666138137673784,
        "ece": 0.016622621027984567,
        "conf": 0.7981700179879287,
        "flip": null,
        "brier": 0.28949881585732284
      },
      {
        "tier": "B",
        "env": "folktables",
        "seed": 1,
        "ce_single": 0.4481469455246447,
        "ce": 0.4356261342600311,
        "diff": {
          "value": -0.012520811264613562,
          "lo": -0.01634915127214725,
          "hi": -0.008893121878611332
        },
        "pass": true,
        "ce_base": 0.6934844593415295,
        "diff_base": {
          "value": -0.25785832508149825,
          "lo": -0.2670034281428999,
          "hi": -0.24913489819862952
        },
        "acc": 0.789375458241685,
        "auroc": 0.7633103515723115,
        "ece": 0.013219735506875569,
        "conf": 0.7976119624434539,
        "flip": null,
        "brier": 0.2863458999243068
      },
      {
        "tier": "B",
        "env": "folktables",
        "seed": 2,
        "ce_single": 0.44233523833101085,
        "ce": 0.4384948531105458,
        "diff": {
          "value": -0.0038403852204651023,
          "lo": -0.007681409785079287,
          "hi": 3.4361367331577135e-05
        },
        "pass": true,
        "ce_base": 0.6934844593415295,
        "diff_base": {
          "value": -0.2549896062309836,
          "lo": -0.26381183562964033,
          "hi": -0.24609928291102195
        },
        "acc": 0.7847097247217223,
        "auroc": 0.7679356316954442,
        "ece": 0.014699667606768206,
        "conf": 0.7962835168970399,
        "flip": null,
        "brier": 0.28842804801350497
      },
      {
        "tier": "B",
        "env": "camelyon17",
        "seed": 0,
        "ce_single": 0.20837835979747724,
        "ce": 0.19917673998548172,
        "diff": {
          "value": -0.009201619811995519,
          "lo": -0.030910606864758623,
          "hi": 0.01116204998378419
        },
        "pass": true,
        "ce_base": 0.6931978275120406,
        "diff_base": {
          "value": -0.4940210875265588,
          "lo": -0.5312363334452189,
          "hi": -0.4495808424622838
        },
        "acc": 0.9294615384615385,
        "auroc": 0.8643750951592453,
        "ece": 0.01728559489892081,
        "conf": 0.9467440238870107,
        "flip": null,
        "brier": 0.1093529491210613
      },
      {
        "tier": "B",
        "env": "camelyon17",
        "seed": 1,
        "ce_single": 0.18892128246007237,
        "ce": 0.2022852801095065,
        "diff": {
          "value": 0.01336399764943412,
          "lo": -0.03134999841335524,
          "hi": 0.052091479403583205
        },
        "pass": true,
        "ce_base": 0.6931978275120406,
        "diff_base": {
          "value": -0.490912547402534,
          "lo": -0.5306599014776814,
          "hi": -0.45281031031847896
        },
        "acc": 0.930076923076923,
        "auroc": 0.8874975786388497,
        "ece": 0.032629293684776,
        "conf": 0.9625870088293003,
        "flip": null,
        "brier": 0.11040756872368296
      },
      {
        "tier": "B",
        "env": "camelyon17",
        "seed": 2,
        "ce_single": 0.22308319161557216,
        "ce": 0.2580159262717116,
        "diff": {
          "value": 0.03493273465613941,
          "lo": 0.01103169983968305,
          "hi": 0.06797831187697921
        },
        "pass": false,
        "ce_base": 0.6931978275120406,
        "diff_base": {
          "value": -0.435181901240329,
          "lo": -0.49185439548196314,
          "hi": -0.3594040908925146
        },
        "acc": 0.9143846153846154,
        "auroc": 0.8487005253347428,
        "ece": 0.03692966717940113,
        "conf": 0.9513100121938265,
        "flip": null,
        "brier": 0.13598931412826326
      }
    ],
    "1.6B": [
      {
        "tier": "B",
        "env": "mcq",
        "seed": 0,
        "ce_single": 0.40774162897171706,
        "ce": 0.4002523221645569,
        "diff": {
          "value": -0.007489306807160107,
          "lo": -0.02144890226507013,
          "hi": 0.00687239297739766
        },
        "pass": true,
        "ce_base": 0.5269270169220264,
        "diff_base": {
          "value": -0.12667469475746948,
          "lo": -0.14512021534410527,
          "hi": -0.10755897094871683
        },
        "acc": 0.8554502369668247,
        "auroc": 0.8781354225007443,
        "ece": 0.0288912705417964,
        "conf": 0.8843396415677963,
        "flip": 0.08807266982622433,
        "brier": 0.20633200495477827
      },
      {
        "tier": "B",
        "env": "mcq",
        "seed": 1,
        "ce_single": 0.39632642582156424,
        "ce": 0.42684009817659985,
        "diff": {
          "value": 0.03051367235503558,
          "lo": 0.01535934910813931,
          "hi": 0.04480322164845855
        },
        "pass": false,
        "ce_base": 0.5269270169220264,
        "diff_base": {
          "value": -0.10008691874542655,
          "lo": -0.11721972238829916,
          "hi": -0.08171584887331773
        },
        "acc": 0.846958925750395,
        "auroc": 0.8766589700584391,
        "ece": 0.012092256853813047,
        "conf": 0.8572383006007703,
        "flip": 0.10624012638230648,
        "brier": 0.21781855156334679
      },
      {
        "tier": "B",
        "env": "mcq",
        "seed": 2,
        "ce_single": 0.4040138206383986,
        "ce": 0.41635328038970587,
        "diff": {
          "value": 0.012339459751307244,
          "lo": -0.003366449172156456,
          "hi": 0.027517208823090694
        },
        "pass": true,
        "ce_base": 0.5269270169220264,
        "diff_base": {
          "value": -0.11057373653232058,
          "lo": -0.13183961328275404,
          "hi": -0.090193696888094
        },
        "acc": 0.8578199052132701,
        "auroc": 0.8752545017393084,
        "ece": 0.0429864913625363,
        "conf": 0.9008063965758064,
        "flip": 0.08708530805687204,
        "brier": 0.2106166671020026
      },
      {
        "tier": "B",
        "env": "chaosnli",
        "seed": 0,
        "ce_single": 0.7432735836870603,
        "ce": 0.7654613715471141,
        "diff": {
          "value": 0.022187787860053886,
          "lo": 0.014892751113375909,
          "hi": 0.029058804683116835
        },
        "pass": false,
        "ce_base": 0.8347258254601914,
        "diff_base": {
          "value": -0.06926445391307712,
          "lo": -0.07996195673824386,
          "hi": -0.058263186156225916
        },
        "acc": 0.7082880926529135,
        "auroc": 0.6902250748410934,
        "ece": 0.030578552813707745,
        "conf": 0.6869863378065374,
        "flip": 0.12015924719507781,
        "brier": 0.1552037163302883
      },
      {
        "tier": "B",
        "env": "chaosnli",
        "seed": 1,
        "ce_single": 0.7425795222297821,
        "ce": 0.7837159885599414,
        "diff": {
          "value": 0.04113646633015925,
          "lo": 0.03226748686400686,
          "hi": 0.04967784705488425
        },
        "pass": false,
        "ce_base": 0.8347258254601914,
        "diff_base": {
          "value": -0.0510098369002501,
          "lo": -0.06193706511579653,
          "hi": -0.0398299530236231
        },
        "acc": 0.6800579080709374,
        "auroc": 0.6697049311393612,
        "ece": 0.02181350362935101,
        "conf": 0.6740838724376929,
        "flip": 0.31921824104234525,
        "brier": 0.17074606097407263
      },
      {
        "tier": "B",
        "env": "chaosnli",
        "seed": 2,
        "ce_single": 0.7426378492384095,
        "ce": 0.7585695789984018,
        "diff": {
          "value": 0.015931729759992495,
          "lo": 0.00928400184774719,
          "hi": 0.023151134788389538
        },
        "pass": false,
        "ce_base": 0.8347258254601914,
        "diff_base": {
          "value": -0.07615624646178952,
          "lo": -0.08783875216411097,
          "hi": -0.06443452143714924
        },
        "acc": 0.7278320665942816,
        "auroc": 0.6720206417893078,
        "ece": 0.056012450127855294,
        "conf": 0.6846000137663043,
        "flip": 0.09554831704668838,
        "brier": 0.14927606706706414
      },
      {
        "tier": "B",
        "env": "cifar10",
        "seed": 0,
        "ce_single": 0.2299739719242439,
        "ce": 0.2519872646207342,
        "diff": {
          "value": 0.022013292696490255,
          "lo": 0.0030736171336475703,
          "hi": 0.03992000338242685
        },
        "pass": false,
        "ce_base": 0.525821451871081,
        "diff_base": {
          "value": -0.27383418725034675,
          "lo": -0.3237711272866988,
          "hi": -0.22284108838891117
        },
        "acc": 0.9315304347826087,
        "auroc": 0.9566540211436751,
        "ece": 0.028470759793986423,
        "conf": 0.9599102325045544,
        "flip": 0.027443478260869566,
        "brier": 0.10188192213687446
      },
      {
        "tier": "B",
        "env": "cifar10",
        "seed": 1,
        "ce_single": 0.2501114400700646,
        "ce": 0.2533726154562614,
        "diff": {
          "value": 0.0032611753861967957,
          "lo": -0.011275453788412921,
          "hi": 0.016673796957155318
        },
        "pass": true,
        "ce_base": 0.525821451871081,
        "diff_base": {
          "value": -0.2724488364148195,
          "lo": -0.32245688367507885,
          "hi": -0.22221876717115202
        },
        "acc": 0.9217565217391305,
        "auroc": 0.9468035615074913,
        "ece": 0.006458782019822494,
        "conf": 0.9263151420883511,
        "flip": 0.04678260869565217,
        "brier": 0.11110962959453999
      },
      {
        "tier": "B",
        "env": "cifar10",
        "seed": 2,
        "ce_single": 0.2321976805761593,
        "ce": 0.334982548927015,
        "diff": {
          "value": 0.10278486835085567,
          "lo": 0.0748897647062431,
          "hi": 0.1358331625647502
        },
        "pass": false,
        "ce_base": 0.525821451871081,
        "diff_base": {
          "value": -0.1908389029440659,
          "lo": -0.23785761084554316,
          "hi": -0.1401684523329021
        },
        "acc": 0.9176521739130434,
        "auroc": 0.936990951591359,
        "ece": 0.04628719613396603,
        "conf": 0.9639267829972765,
        "flip": 0.03248695652173913,
        "brier": 0.1294624052588854
      },
      {
        "tier": "B",
        "env": "civilcomments",
        "seed": 0,
        "ce_single": 0.31429617484808725,
        "ce": 0.3291432395939985,
        "diff": {
          "value": 0.014847064745911189,
          "lo": 0.013104040172484954,
          "hi": 0.01664621848809903
        },
        "pass": false,
        "ce_base": 0.6286493075695474,
        "diff_base": {
          "value": -0.299506067975549,
          "lo": -0.3074177196746435,
          "hi": -0.2915985168841146
        },
        "acc": 0.9314809564313392,
        "auroc": 0.9217205693013882,
        "ece": 0.04325364368555168,
        "conf": 0.8949943995861078,
        "flip": null,
        "brier": 0.04824225276040297
      },
      {
        "tier": "B",
        "env": "civilcomments",
        "seed": 1,
        "ce_single": 0.3138288534673895,
        "ce": 0.3231331506493501,
        "diff": {
          "value": 0.009304297181960621,
          "lo": 0.007857961523336704,
          "hi": 0.010786769292632555
        },
        "pass": false,
        "ce_base": 0.6286493075695474,
        "diff_base": {
          "value": -0.30551615692019735,
          "lo": -0.31285911893489743,
          "hi": -0.2983232581327826
        },
        "acc": 0.93346564596919,
        "auroc": 0.9217670673143299,
        "ece": 0.0636185640232424,
        "conf": 0.8726835281555306,
        "flip": null,
        "brier": 0.04397100215369724
      },
      {
        "tier": "B",
        "env": "civilcomments",
        "seed": 2,
        "ce_single": 0.3146981139398467,
        "ce": 0.3163261854007762,
        "diff": {
          "value": 0.00162807146092949,
          "lo": 0.0005988585567799004,
          "hi": 0.002661235833755524
        },
        "pass": false,
        "ce_base": 0.6286493075695474,
        "diff_base": {
          "value": -0.31232312216877123,
          "lo": -0.31959109428057375,
          "hi": -0.3052725327859585
        },
        "acc": 0.9425385124279368,
        "auroc": 0.9198111216072871,
        "ece": 0.07780369303648059,
        "conf": 0.8647348193914562,
        "flip": null,
        "brier": 0.039377961884232315
      },
      {
        "tier": "B",
        "env": "stsb",
        "seed": 0,
        "ce_single": 1.055144585762775,
        "ce": 1.031748552907778,
        "diff": {
          "value": -0.023396032854997285,
          "lo": -0.04090748294174989,
          "hi": -0.0036253573610600352
        },
        "pass": true,
        "ce_base": 1.749101240806852,
        "diff_base": {
          "value": -0.7173526878990744,
          "lo": -0.8007907469966399,
          "hi": -0.6449286724678583
        },
        "acc": 0.6131664853101197,
        "auroc": 0.6128823020677726,
        "ece": 0.052070733139512326,
        "conf": 0.6504272062481422,
        "flip": null,
        "brier": 0.290818102811832
      },
      {
        "tier": "B",
        "env": "stsb",
        "seed": 1,
        "ce_single": 1.0539440524395638,
        "ce": 1.0325978650062761,
        "diff": {
          "value": -0.021346187433287877,
          "lo": -0.05153571897527513,
          "hi": 0.008768368946593146
        },
        "pass": true,
        "ce_base": 1.749101240806852,
        "diff_base": {
          "value": -0.716503375800576,
          "lo": -0.7897844611529958,
          "hi": -0.6517002411893676
        },
        "acc": 0.5715451577801959,
        "auroc": 0.621027477467268,
        "ece": 0.04524248022022652,
        "conf": 0.6035776622552866,
        "flip": null,
        "brier": 0.30203148903766214
      },
      {
        "tier": "B",
        "env": "stsb",
        "seed": 2,
        "ce_single": 1.0425603202621663,
        "ce": 1.1350048984988437,
        "diff": {
          "value": 0.09244457823667729,
          "lo": 0.05544914470844475,
          "hi": 0.13377877956106016
        },
        "pass": false,
        "ce_base": 1.749101240806852,
        "diff_base": {
          "value": -0.6140963423080084,
          "lo": -0.7119540303207893,
          "hi": -0.525878957121933
        },
        "acc": 0.5998367791077258,
        "auroc": 0.6290156941997284,
        "ece": 0.12202087906518622,
        "conf": 0.721857658172912,
        "flip": null,
        "brier": 0.3191667956646663
      },
      {
        "tier": "B",
        "env": "folktables",
        "seed": 0,
        "ce_single": 0.43414523920300574,
        "ce": 0.4429014798100703,
        "diff": {
          "value": 0.008756240607064609,
          "lo": 0.005746677057294381,
          "hi": 0.011726213381559743
        },
        "pass": false,
        "ce_base": 0.6121096088424054,
        "diff_base": {
          "value": -0.1692081290323351,
          "lo": -0.1793469097741068,
          "hi": -0.15912578607336217
        },
        "acc": 0.7856428714257149,
        "auroc": 0.760233718887534,
        "ece": 0.025429919649106476,
        "conf": 0.8061644708856792,
        "flip": null,
        "brier": 0.29192483263490315
      },
      {
        "tier": "B",
        "env": "folktables",
        "seed": 1,
        "ce_single": 0.4378669224129474,
        "ce": 0.44795320351265,
        "diff": {
          "value": 0.010086281099702563,
          "lo": 0.0053899754819800244,
          "hi": 0.014885390200504426
        },
        "pass": false,
        "ce_base": 0.6121096088424054,
        "diff_base": {
          "value": -0.16415640532975542,
          "lo": -0.17344494183186426,
          "hi": -0.15514808425113147
        },
        "acc": 0.7784443111377725,
        "auroc": 0.760897827041841,
        "ece": 0.013490881572951727,
        "conf": 0.7826558687977256,
        "flip": null,
        "brier": 0.29621629260680155
      },
      {
        "tier": "B",
        "env": "folktables",
        "seed": 2,
        "ce_single": 0.43541280816255606,
        "ce": 0.4356590369299968,
        "diff": {
          "value": 0.0002462287674407353,
          "lo": -0.002834348848603956,
          "hi": 0.0033126552922933578
        },
        "pass": true,
        "ce_base": 0.6121096088424054,
        "diff_base": {
          "value": -0.1764505719124086,
          "lo": -0.1862956406214327,
          "hi": -0.16675564239316432
        },
        "acc": 0.7911084449776711,
        "auroc": 0.7657509738706254,
        "ece": 0.01979037306272671,
        "conf": 0.8088698498306295,
        "flip": null,
        "brier": 0.28450758265424486
      },
      {
        "tier": "B",
        "env": "camelyon17",
        "seed": 0,
        "ce_single": 0.14715429696058982,
        "ce": 0.16948480361084028,
        "diff": {
          "value": 0.022330506650250492,
          "lo": 0.005246748162873034,
          "hi": 0.0401075197620554
        },
        "pass": false,
        "ce_base": 0.6492519817213229,
        "diff_base": {
          "value": -0.4797671781104826,
          "lo": -0.5181835922538937,
          "hi": -0.4500954663723073
        },
        "acc": 0.9452307692307692,
        "auroc": 0.8774072311344218,
        "ece": 0.024357909413484414,
        "conf": 0.9695761512517929,
        "flip": null,
        "brier": 0.08766813727769751
      },
      {
        "tier": "B",
        "env": "camelyon17",
        "seed": 1,
        "ce_single": 0.15899906997961716,
        "ce": 0.15577225450669627,
        "diff": {
          "value": -0.0032268154729209056,
          "lo": -0.02846167486233926,
          "hi": 0.018547685804321406
        },
        "pass": true,
        "ce_base": 0.6492519817213229,
        "diff_base": {
          "value": -0.4934797272146266,
          "lo": -0.5387851743221831,
          "hi": -0.4622773848595618
        },
        "acc": 0.9384615384615385,
        "auroc": 0.907457530737705,
        "ece": 0.013398746646367605,
        "conf": 0.9395801230302224,
        "flip": null,
        "brier": 0.0900247205040645
      },
      {
        "tier": "B",
        "env": "camelyon17",
        "seed": 2,
        "ce_single": 0.1814494749853387,
        "ce": 0.16058040959806322,
        "diff": {
          "value": -0.020869065387275485,
          "lo": -0.033328658363837894,
          "hi": -0.009703955758107462
        },
        "pass": true,
        "ce_base": 0.6492519817213229,
        "diff_base": {
          "value": -0.4886715721232597,
          "lo": -0.5259585444754835,
          "hi": -0.45916087180637793
        },
        "acc": 0.941923076923077,
        "auroc": 0.8967018839964414,
        "ece": 0.020103888387863443,
        "conf": 0.9619018222047733,
        "flip": null,
        "brier": 0.09038797778880224
      }
    ],
    "3B": [
      {
        "tier": "B",
        "env": "mcq",
        "seed": 0,
        "ce": 0.28394581130659097,
        "ce_base": 0.38602727384245317,
        "diff_base": {
          "value": -0.10208146253586213,
          "lo": -0.11673248732151378,
          "hi": -0.08761654927776354
        },
        "acc": 0.8992890995260664,
        "auroc": 0.9004529523711109,
        "ece": 0.01118656555058266,
        "conf": 0.9096089693560231,
        "brier": 0.14707832873591406
      },
      {
        "tier": "B",
        "env": "mcq",
        "seed": 1,
        "ce": 0.2891957645267851,
        "ce_base": 0.38602727384245317,
        "diff_base": {
          "value": -0.09683150931566804,
          "lo": -0.11046495538913444,
          "hi": -0.08322018033501698
        },
        "acc": 0.8937598736176935,
        "auroc": 0.9045484002385228,
        "ece": 0.00769100147542233,
        "conf": 0.8991462239086345,
        "brier": 0.1517139060824241
      },
      {
        "tier": "B",
        "env": "mcq",
        "seed": 2,
        "ce": 0.2861016989829702,
        "ce_base": 0.38602727384245317,
        "diff_base": {
          "value": -0.09992557485948292,
          "lo": -0.11570205023000027,
          "hi": -0.08457066767730967
        },
        "acc": 0.8996840442338072,
        "auroc": 0.9013879335375001,
        "ece": 0.02074630266588068,
        "conf": 0.9201619169346718,
        "brier": 0.1476629415774208
      },
      {
        "tier": "B",
        "env": "chaosnli",
        "seed": 0,
        "ce": 0.7317921943962872,
        "ce_base": 0.7949680342778785,
        "diff_base": {
          "value": -0.06317583988159142,
          "lo": -0.07211168590286045,
          "hi": -0.05446602001900818
        },
        "acc": 0.7321751719145856,
        "auroc": 0.7111210939065611,
        "ece": 0.048137297006435174,
        "conf": 0.6842451099248369,
        "brier": 0.13177767445592195
      },
      {
        "tier": "B",
        "env": "chaosnli",
        "seed": 1,
        "ce": 0.7349974377531698,
        "ce_base": 0.7949680342778785,
        "diff_base": {
          "value": -0.059970596524708736,
          "lo": -0.06850646719340128,
          "hi": -0.05143582608454075
        },
        "acc": 0.7285559174809989,
        "auroc": 0.7173164431197218,
        "ece": 0.04855649530822884,
        "conf": 0.687175090378538,
        "brier": 0.13334485416438893
      },
      {
        "tier": "B",
        "env": "chaosnli",
        "seed": 2,
        "ce": 0.7314480272773598,
        "ce_base": 0.7949680342778785,
        "diff_base": {
          "value": -0.06352000700051882,
          "lo": -0.0721910035084413,
          "hi": -0.054702471241274524
        },
        "acc": 0.7350705754614549,
        "auroc": 0.698795715588703,
        "ece": 0.054777696259505855,
        "conf": 0.6815311866871395,
        "brier": 0.13283454092153876
      },
      {
        "tier": "B",
        "env": "cifar10",
        "seed": 0,
        "ce": 0.1926190952540813,
        "ce_base": 0.2611563624941145,
        "diff_base": {
          "value": -0.06853726724003315,
          "lo": -0.09402452041400289,
          "hi": -0.04432714693000657
        },
        "acc": 0.9390434782608695,
        "auroc": 0.9529923188587097,
        "ece": 0.016654349529743186,
        "conf": 0.9556232028479161,
        "brier": 0.08987807640548674
      },
      {
        "tier": "B",
        "env": "cifar10",
        "seed": 1,
        "ce": 0.19700957925621565,
        "ce_base": 0.2611563624941145,
        "diff_base": {
          "value": -0.06414678323789884,
          "lo": -0.08776735215450746,
          "hi": -0.041930136281348514
        },
        "acc": 0.9375652173913044,
        "auroc": 0.9508026273025971,
        "ece": 0.01768210469406584,
        "conf": 0.9533248674587064,
        "brier": 0.0916005966374745
      },
      {
        "tier": "B",
        "env": "cifar10",
        "seed": 2,
        "ce": 0.24005205260956894,
        "ce_base": 0.2611563624941145,
        "diff_base": {
          "value": -0.02110430988454551,
          "lo": -0.04498645729999961,
          "hi": 0.0030856171380235138
        },
        "acc": 0.9355478260869565,
        "auroc": 0.9478078197050045,
        "ece": 0.029083070768221565,
        "conf": 0.9646026096185912,
        "brier": 0.09852223326178403
      },
      {
        "tier": "B",
        "env": "civilcomments",
        "seed": 0,
        "ce": 0.32493529455042947,
        "ce_base": 0.422683145843368,
        "diff_base": {
          "value": -0.09774785129293849,
          "lo": -0.10272723555639415,
          "hi": -0.09211969004446136
        },
        "acc": 0.9376240430961157,
        "auroc": 0.9210726863433244,
        "ece": 0.045806885644400594,
        "conf": 0.8938100220361445,
        "brier": 0.044440062724136684
      },
      {
        "tier": "B",
        "env": "civilcomments",
        "seed": 1,
        "ce": 0.31942936538814876,
        "ce_base": 0.422683145843368,
        "diff_base": {
          "value": -0.10325378045521923,
          "lo": -0.10829973730458461,
          "hi": -0.09770082750583385
        },
        "acc": 0.9445232019657878,
        "auroc": 0.913231925526696,
        "ece": 0.06654398142305668,
        "conf": 0.877979220542731,
        "brier": 0.040234408557455296
      },
      {
        "tier": "B",
        "env": "civilcomments",
        "seed": 2,
        "ce": 0.3162963367925811,
        "ce_base": 0.422683145843368,
        "diff_base": {
          "value": -0.1063868090507868,
          "lo": -0.1111116179990059,
          "hi": -0.1012086043872098
        },
        "acc": 0.9453737831962953,
        "auroc": 0.9157116186943225,
        "ece": 0.08202986666074875,
        "conf": 0.8633439165355465,
        "brier": 0.0392586704145921
      },
      {
        "tier": "B",
        "env": "stsb",
        "seed": 0,
        "ce": 1.0320934313543655,
        "ce_base": 1.5758103744712262,
        "diff_base": {
          "value": -0.5437169431168607,
          "lo": -0.6207747093223118,
          "hi": -0.4716511781174488
        },
        "acc": 0.6235038084874864,
        "auroc": 0.6192176040815504,
        "ece": 0.06937201697302331,
        "conf": 0.6865655293564542,
        "brier": 0.2879302523060823
      },
      {
        "tier": "B",
        "env": "stsb",
        "seed": 1,
        "ce": 0.9606397550106914,
        "ce_base": 1.5758103744712262,
        "diff_base": {
          "value": -0.6151706194605348,
          "lo": -0.6804369290481779,
          "hi": -0.5581021087930027
        },
        "acc": 0.6109902067464635,
        "auroc": 0.6517820336386676,
        "ece": 0.03286231362683471,
        "conf": 0.6308347471342123,
        "brier": 0.2696511629637947
      },
      {
        "tier": "B",
        "env": "stsb",
        "seed": 2,
        "ce": 1.0124180703390282,
        "ce_base": 1.5758103744712262,
        "diff_base": {
          "value": -0.563392304132198,
          "lo": -0.6365432112290809,
          "hi": -0.49533936739389856
        },
        "acc": 0.6191512513601741,
        "auroc": 0.6282491212653779,
        "ece": 0.0637306943779023,
        "conf": 0.6741650431245663,
        "brier": 0.28177827479458084
      },
      {
        "tier": "B",
        "env": "folktables",
        "seed": 0,
        "ce": 0.44509115726019133,
        "ce_base": 0.609676541729926,
        "diff_base": {
          "value": -0.16458538446973475,
          "lo": -0.17439749195793422,
          "hi": -0.15483688868750595
        },
        "acc": 0.7876424715056989,
        "auroc": 0.7532512981367189,
        "ece": 0.02779868638635452,
        "conf": 0.8070413636770326,
        "brier": 0.2930003292994701
      },
      {
        "tier": "B",
        "env": "folktables",
        "seed": 1,
        "ce": 0.44111304791622236,
        "ce_base": 0.609676541729926,
        "diff_base": {
          "value": -0.16856349381370364,
          "lo": -0.17746318378000042,
          "hi": -0.16005416847899803
        },
        "acc": 0.7869759381457042,
        "auroc": 0.7613656299221321,
        "ece": 0.014096905742098072,
        "conf": 0.7849255870844709,
        "brier": 0.289130483213222
      },
      {
        "tier": "B",
        "env": "folktables",
        "seed": 2,
        "ce": 0.45390238176418163,
        "ce_base": 0.609676541729926,
        "diff_base": {
          "value": -0.15577415996574437,
          "lo": -0.16661552004733296,
          "hi": -0.1447391492070872
        },
        "acc": 0.7881090448576952,
        "auroc": 0.7578438227483217,
        "ece": 0.03934656827201734,
        "conf": 0.8258664168088204,
        "brier": 0.2935044644711353
      },
      {
        "tier": "B",
        "env": "camelyon17",
        "seed": 0,
        "ce": 0.1406058934525335,
        "ce_base": 0.6861919247772764,
        "diff_base": {
          "value": -0.5455860313247429,
          "lo": -0.6275696224676554,
          "hi": -0.4919595356019622
        },
        "acc": 0.9468461538461539,
        "auroc": 0.9067836424796653,
        "ece": 0.011683198690414417,
        "conf": 0.9584915467592386,
        "brier": 0.07937519631993027
      },
      {
        "tier": "B",
        "env": "camelyon17",
        "seed": 1,
        "ce": 0.19288188233974013,
        "ce_base": 0.6861919247772764,
        "diff_base": {
          "value": -0.4933100424375363,
          "lo": -0.5776901705796982,
          "hi": -0.43618574916489894
        },
        "acc": 0.942076923076923,
        "auroc": 0.8713108698544598,
        "ece": 0.02991748259159234,
        "conf": 0.9719926106517132,
        "brier": 0.09324804513698431
      },
      {
        "tier": "B",
        "env": "camelyon17",
        "seed": 2,
        "ce": 0.1634219955634534,
        "ce_base": 0.6861919247772764,
        "diff_base": {
          "value": -0.522769929213823,
          "lo": -0.5845630301942749,
          "hi": -0.48434549592554427
        },
        "acc": 0.938076923076923,
        "auroc": 0.8971406670588445,
        "ece": 0.012466411677690642,
        "conf": 0.9502902042086308,
        "brier": 0.09485444071675608
      }
    ]
  },
  "missing_values": "Non-finite values in the source reports (e.g. unmeasured flip rates) are represented as null."
}
