{
  "temperature": 1.0087398771573446,
  "fitted_on": "calibration",
  "selected_on": "validation",
  "use_temperature": false,
  "default_prediction": "raw",
  "validation_raw": {
    "count": 14111,
    "top1_accuracy": 0.8263765856424067,
    "nll": 0.6840500261181026,
    "brier_score": 0.011582265527708373,
    "ece": 0.16699567031458232,
    "signed_calibration_error": -0.1669956703145823,
    "kl_divergence": 0.02310311772478613,
    "confidence_analysis": {
      "correct": {
        "count": 11661,
        "mean": 0.6960411451404903,
        "median": 0.6765217781066895,
        "std": 0.19690960502309285,
        "p25": 0.50345379114151,
        "p75": 0.8713327646255493
      },
      "incorrect": {
        "count": 2450,
        "mean": 0.4848931847786417,
        "median": 0.500392496585846,
        "std": 0.07416498496851862,
        "p25": 0.5000284314155579,
        "p75": 0.5010085254907608
      }
    },
    "confidence_histogram": [
      {
        "bucket": "0.0-0.1",
        "sample_count": 0,
        "average_confidence": null,
        "empirical_accuracy": null,
        "signed_gap": null
      },
      {
        "bucket": "0.1-0.2",
        "sample_count": 0,
        "average_confidence": null,
        "empirical_accuracy": null,
        "signed_gap": null
      },
      {
        "bucket": "0.2-0.3",
        "sample_count": 148,
        "average_confidence": 0.27538005906987834,
        "empirical_accuracy": 0.46621621621621623,
        "signed_gap": -0.1908361571463379
      },
      {
        "bucket": "0.3-0.4",
        "sample_count": 667,
        "average_confidence": 0.3564207344755776,
        "empirical_accuracy": 0.664167916041979,
        "signed_gap": -0.3077471815664014
      },
      {
        "bucket": "0.4-0.5",
        "sample_count": 1160,
        "average_confidence": 0.4550788116095395,
        "empirical_accuracy": 0.7741379310344828,
        "signed_gap": -0.3190591194249433
      },
      {
        "bucket": "0.5-0.6",
        "sample_count": 4972,
        "average_confidence": 0.5200908452703985,
        "empirical_accuracy": 0.6351568785197104,
        "signed_gap": -0.11506603324931186
      },
      {
        "bucket": "0.6-0.7",
        "sample_count": 1678,
        "average_confidence": 0.6489259644685684,
        "empirical_accuracy": 0.9713945172824792,
        "signed_gap": -0.32246855281391074
      },
      {
        "bucket": "0.7-0.8",
        "sample_count": 1540,
        "average_confidence": 0.7498124199641215,
        "empirical_accuracy": 0.9935064935064936,
        "signed_gap": -0.24369407354237205
      },
      {
        "bucket": "0.8-0.9",
        "sample_count": 1412,
        "average_confidence": 0.8490157142154894,
        "empirical_accuracy": 0.9936260623229461,
        "signed_gap": -0.14461034810745677
      },
      {
        "bucket": "0.9-1.0",
        "sample_count": 2534,
        "average_confidence": 0.9746775966087663,
        "empirical_accuracy": 0.9984214680347278,
        "signed_gap": -0.02374387142596146
      }
    ],
    "selective_accuracy": [
      {
        "threshold": 0.5,
        "coverage": 0.8600382680178584,
        "accuracy": 0.844676994067238,
        "risk": 0.15532300593276205,
        "sample_count": 12136
      },
      {
        "threshold": 0.6,
        "coverage": 0.5076890369215505,
        "accuracy": 0.9900893355667225,
        "risk": 0.009910664433277483,
        "sample_count": 7164
      },
      {
        "threshold": 0.7,
        "coverage": 0.3887747147615336,
        "accuracy": 0.9958075100255195,
        "risk": 0.004192489974480518,
        "sample_count": 5486
      },
      {
        "threshold": 0.8,
        "coverage": 0.279639997165332,
        "accuracy": 0.996705524581855,
        "risk": 0.003294475418144982,
        "sample_count": 3946
      },
      {
        "threshold": 0.9,
        "coverage": 0.17957621713556798,
        "accuracy": 0.9984214680347278,
        "risk": 0.0015785319652722452,
        "sample_count": 2534
      },
      {
        "threshold": 0.95,
        "coverage": 0.13273332860888667,
        "accuracy": 0.9983982915109451,
        "risk": 0.0016017084890549427,
        "sample_count": 1873
      }
    ]
  },
  "validation_temperature_scaled": {
    "count": 14111,
    "top1_accuracy": 0.8263057189426688,
    "nll": 0.6840468913421749,
    "brier_score": 0.011582336771387319,
    "ece": 0.16813608304214336,
    "signed_calibration_error": -0.16813608304214334,
    "kl_divergence": 0.023099982948858303,
    "confidence_analysis": {
      "correct": {
        "count": 11660,
        "mean": 0.694681208951386,
        "median": 0.6747364841480057,
        "std": 0.19693845421787254,
        "p25": 0.5021719095333224,
        "p75": 0.8689303683324948
      },
      "incorrect": {
        "count": 2451,
        "mean": 0.4844752492122222,
        "median": 0.5003890293824186,
        "std": 0.07420726284342961,
        "p25": 0.500027128879432,
        "p75": 0.5009866390613054
      }
    },
    "confidence_histogram": [
      {
        "bucket": "0.0-0.1",
        "sample_count": 0,
        "average_confidence": null,
        "empirical_accuracy": null,
        "signed_gap": null
      },
      {
        "bucket": "0.1-0.2",
        "sample_count": 0,
        "average_confidence": null,
        "empirical_accuracy": null,
        "signed_gap": null
      },
      {
        "bucket": "0.2-0.3",
        "sample_count": 151,
        "average_confidence": 0.2751785209954829,
        "empirical_accuracy": 0.46357615894039733,
        "signed_gap": -0.18839763794491443
      },
      {
        "bucket": "0.3-0.4",
        "sample_count": 674,
        "average_confidence": 0.35612814661984515,
        "empirical_accuracy": 0.6646884272997032,
        "signed_gap": -0.3085602806798581
      },
      {
        "bucket": "0.4-0.5",
        "sample_count": 1168,
        "average_confidence": 0.45463179276212934,
        "empirical_accuracy": 0.7782534246575342,
        "signed_gap": -0.3236216318954049
      },
      {
        "bucket": "0.5-0.6",
        "sample_count": 4987,
        "average_confidence": 0.5201110232818913,
        "empirical_accuracy": 0.6360537397232805,
        "signed_gap": -0.1159427164413892
      },
      {
        "bucket": "0.6-0.7",
        "sample_count": 1673,
        "average_confidence": 0.6487813857971743,
        "empirical_accuracy": 0.9719067543335326,
        "signed_gap": -0.3231253685363583
      },
      {
        "bucket": "0.7-0.8",
        "sample_count": 1546,
        "average_confidence": 0.7495579579669318,
        "empirical_accuracy": 0.9935316946959897,
        "signed_gap": -0.24397373672905787
      },
      {
        "bucket": "0.8-0.9",
        "sample_count": 1409,
        "average_confidence": 0.8491268025105129,
        "empirical_accuracy": 0.99361249112846,
        "signed_gap": -0.14448568861794708
      },
      {
        "bucket": "0.9-1.0",
        "sample_count": 2503,
        "average_confidence": 0.9749893213960854,
        "empirical_accuracy": 0.9984019176987615,
        "signed_gap": -0.023412596302676114
      }
    ],
    "selective_accuracy": [
      {
        "threshold": 0.5,
        "coverage": 0.8587626674225781,
        "accuracy": 0.8444462782637399,
        "risk": 0.15555372173626014,
        "sample_count": 12118
      },
      {
        "threshold": 0.6,
        "coverage": 0.5053504358302033,
        "accuracy": 0.9901837049502173,
        "risk": 0.009816295049782653,
        "sample_count": 7131
      },
      {
        "threshold": 0.7,
        "coverage": 0.38679044716887534,
        "accuracy": 0.9957860021986076,
        "risk": 0.004213997801392422,
        "sample_count": 5458
      },
      {
        "threshold": 0.8,
        "coverage": 0.277230529374247,
        "accuracy": 0.996676891615542,
        "risk": 0.00332310838445804,
        "sample_count": 3912
      },
      {
        "threshold": 0.9,
        "coverage": 0.1773793494436964,
        "accuracy": 0.9984019176987615,
        "risk": 0.0015980823012384793,
        "sample_count": 2503
      },
      {
        "threshold": 0.95,
        "coverage": 0.131599461413082,
        "accuracy": 0.9983844911147012,
        "risk": 0.001615508885298822,
        "sample_count": 1857
      }
    ]
  }
}