calibration.json

{
  "a": -1.8895045041017524,
  "b1": 0.48729786088254085,
  "b2": 0.5583580780553591,
  "fit_split": "data/dev.jsonl",
  "n_fit": 2786,
  "exact_las_fit": 0.8406,
  "ece_raw_fit": 0.1255,
  "ece_cal_fit": 0.0102,
  "table_fit": [
    {
      "threshold": 0.5,
      "emitted": 2290,
      "emitted_pct": 0.822,
      "accuracy": 0.959
    },
    {
      "threshold": 0.6,
      "emitted": 2219,
      "emitted_pct": 0.7965,
      "accuracy": 0.9721
    },
    {
      "threshold": 0.7,
      "emitted": 2147,
      "emitted_pct": 0.7706,
      "accuracy": 0.9818
    },
    {
      "threshold": 0.8,
      "emitted": 2087,
      "emitted_pct": 0.7491,
      "accuracy": 0.989
    },
    {
      "threshold": 0.9,
      "emitted": 1994,
      "emitted_pct": 0.7157,
      "accuracy": 0.9935
    },
    {
      "threshold": 0.95,
      "emitted": 1922,
      "emitted_pct": 0.6899,
      "accuracy": 0.9969
    },
    {
      "threshold": 0.98,
      "emitted": 1865,
      "emitted_pct": 0.6694,
      "accuracy": 0.9995
    },
    {
      "threshold": 0.99,
      "emitted": 1827,
      "emitted_pct": 0.6558,
      "accuracy": 1.0
    }
  ],
  "topk_fit": [
    {
      "top_pct": 5,
      "emitted": 139,
      "accuracy": 1.0
    },
    {
      "top_pct": 10,
      "emitted": 278,
      "accuracy": 1.0
    },
    {
      "top_pct": 25,
      "emitted": 696,
      "accuracy": 1.0
    },
    {
      "top_pct": 50,
      "emitted": 1393,
      "accuracy": 1.0
    }
  ],
  "eval_split": "data/test.jsonl",
  "n_eval": 2788,
  "exact_las_eval": 0.8275,
  "ece_raw_eval": 0.1354,
  "ece_cal_eval": 0.0106,
  "table_eval": [
    {
      "threshold": 0.5,
      "emitted": 2256,
      "emitted_pct": 0.8092,
      "accuracy": 0.9605
    },
    {
      "threshold": 0.6,
      "emitted": 2195,
      "emitted_pct": 0.7873,
      "accuracy": 0.9745
    },
    {
      "threshold": 0.7,
      "emitted": 2133,
      "emitted_pct": 0.7651,
      "accuracy": 0.9841
    },
    {
      "threshold": 0.8,
      "emitted": 2083,
      "emitted_pct": 0.7471,
      "accuracy": 0.9899
    },
    {
      "threshold": 0.9,
      "emitted": 1995,
      "emitted_pct": 0.7156,
      "accuracy": 0.9955
    },
    {
      "threshold": 0.95,
      "emitted": 1925,
      "emitted_pct": 0.6905,
      "accuracy": 0.9974
    },
    {
      "threshold": 0.98,
      "emitted": 1868,
      "emitted_pct": 0.67,
      "accuracy": 0.9995
    },
    {
      "threshold": 0.99,
      "emitted": 1837,
      "emitted_pct": 0.6589,
      "accuracy": 0.9995
    }
  ],
  "topk_eval": [
    {
      "top_pct": 5,
      "emitted": 139,
      "accuracy": 1.0
    },
    {
      "top_pct": 10,
      "emitted": 278,
      "accuracy": 1.0
    },
    {
      "top_pct": 25,
      "emitted": 697,
      "accuracy": 1.0
    },
    {
      "top_pct": 50,
      "emitted": 1394,
      "accuracy": 1.0
    }
  ]
}