kodama-core / calibration.json
Cem13's picture
Release Kodama Core with weights, runtime, attribution, and evaluation
c7893fa verified
Raw History Blame Contribute Delete
2.51 kB
{
"buckets": [
[
2,
2,
"2"
],
[
3,
5,
"3-5"
],
[
6,
10,
"6-10"
],
[
11,
30,
"11-30"
],
[
31,
1000000000,
"31+"
]
],
"temperatures": {
"choice": {
"2": 1.3510327339172363,
"3-5": 1.2677322626113892,
"6-10": 1.2146227359771729,
"11-30": 1.010597825050354,
"31+": 0.9884432554244995
},
"score": {
"2": 1.3455543518066406,
"3-5": 1.3487545251846313,
"6-10": 1.3455543518066406,
"11-30": 1.3455543518066406,
"31+": 1.3455543518066406
},
"noul": {
"2": 1.4132403135299683,
"3-5": 1.4132403135299683,
"6-10": 1.4132403135299683,
"11-30": 1.4132403135299683,
"31+": 1.4132403135299683
}
},
"type_temperature": {
"choice": 1.1966118812561035,
"score": 1.3455543518066406,
"noul": 1.4132403135299683
},
"escalate": {
"a": 0.8075713515281677,
"b": -0.10456699132919312
},
"counts": {
"choice": {
"2": 1440,
"3-5": 3550,
"6-10": 1019,
"11-30": 656,
"31+": 225
},
"score": {
"2": 35,
"3-5": 2354,
"6-10": 12,
"11-30": 0,
"31+": 0
},
"noul": {
"2": 6229,
"3-5": 0,
"6-10": 0,
"11-30": 0,
"31+": 0
}
},
"report": {
"val_cal.jsonl": {
"before": {
"n": 15520,
"nll": 0.6214327216148376,
"acc": 0.7330541014671326,
"ece": 0.03956293314695358,
"esc_bce": 0.4619186222553253,
"choice/n": 6890,
"choice/acc": 0.7416545748710632,
"choice/nll": 0.6412952542304993,
"choice/ece": 0.0349237397313118,
"score/n": 2401,
"score/acc": 0.6014160513877869,
"score/nll": 0.9590876698493958,
"score/ece": 0.06188902631402016,
"noul/n": 6229,
"noul/acc": 0.7742815613746643,
"noul/nll": 0.46931150555610657,
"noul/ece": 0.04082968086004257,
"score/rps": 0.13425556469803043,
"macro_acc": 0.7057840625445048
},
"after": {
"n": 15520,
"nll": 0.6076826453208923,
"acc": 0.7330541014671326,
"ece": 0.008992322720587254,
"esc_bce": 0.4538232982158661,
"choice/n": 6890,
"choice/acc": 0.7416545748710632,
"choice/nll": 0.6310363411903381,
"choice/ece": 0.015944216400384903,
"score/n": 2401,
"score/acc": 0.6014160513877869,
"score/nll": 0.9385183453559875,
"score/ece": 0.03042321652173996,
"noul/n": 6229,
"noul/acc": 0.7742815613746643,
"noul/nll": 0.454328328371048,
"noul/ece": 0.01998109556734562,
"score/rps": 0.13217337983058516,
"macro_acc": 0.7057840625445048
}
}
}
}