ZTC-Judge-27B / ztc_config.json
SeaWolf-AI's picture
config: corrected per-domain LODO (curved probe), leaderboard 0.7282, baseline 0.6223
7491b79 verified
Raw History Blame Contribute Delete
2.92 kB
{
"ztc_version": "1.0",
"base_model": "Qwen/Qwen3.5-27B",
"base_revision": "fc05daec18b0a78c049392ed2e771dde82bdf654",
"probe_file": "ZTC-Judge-27B.npz",
"hidden_dim": 5120,
"layer": "last",
"token": "last_non_pad",
"max_length": 2048,
"template": "다음은 어떤 문제와 그에 대한 답변이다. 이 답변이 옳은지 판단하라.\n\n[문제]\n%s\n\n[답변]\n%s\n\n이 답변은 옳은가?",
"thresholds": {
"review": -0.392403355096007,
"clear": 0.11326689029648167
},
"measured": {
"cv_auc": 0.7999247390498649,
"permutation_z": 25.68107120241526,
"n": 2018,
"n_wrong": 508,
"per_domain_lodo": {
"전문시험(법·수학·생물)": 0.8462,
"상식·다단계추론": 0.6172,
"생물·의학": 0.7154,
"과학추론": 0.741,
"재난·안전 지침": 0.6961
},
"per_domain_detail": {
"전문시험(법·수학·생물)": {
"auc": 0.8462,
"z": 9.07,
"n": 400,
"n_wrong": 80
},
"상식·다단계추론": {
"auc": 0.6172,
"z": 3.63,
"n": 278,
"n_wrong": 130
},
"생물·의학": {
"auc": 0.7154,
"z": 8.92,
"n": 917,
"n_wrong": 169
},
"과학추론": {
"auc": 0.741,
"z": 3.82,
"n": 198,
"n_wrong": 20
},
"재난·안전 지침": {
"auc": 0.6961,
"z": 4.72,
"n": 225,
"n_wrong": 109
}
},
"review_budget": [
{
"frac": 0.05,
"k": 100,
"caught": 71,
"recall": 0.13976377952755906,
"precision": 0.71
},
{
"frac": 0.1,
"k": 201,
"caught": 142,
"recall": 0.2795275590551181,
"precision": 0.7064676616915423
},
{
"frac": 0.15,
"k": 302,
"caught": 195,
"recall": 0.3838582677165354,
"precision": 0.6456953642384106
},
{
"frac": 0.2,
"k": 403,
"caught": 240,
"recall": 0.47244094488188976,
"precision": 0.5955334987593052
},
{
"frac": 0.3,
"k": 605,
"caught": 315,
"recall": 0.6200787401574803,
"precision": 0.5206611570247934
},
{
"frac": 0.4,
"k": 807,
"caught": 369,
"recall": 0.7263779527559056,
"precision": 0.45724907063197023
}
],
"leaderboard_auc": 0.7282,
"surface_baseline": 0.6223,
"protocol": "2,018 items, corrected set; leave-one-domain-out; hyper-parameters chosen inside training domains only; per-domain then size-weighted. Curved probe (ztc_curve_probe.npz).",
"note_linear": "cv_auc 0.7999 comes from the linear probe on 400 professional-exam items with random folds — an easier protocol. Quote leaderboard_auc when comparing systems."
},
"note": "per_domain_lodo 는 leave-one-domain-out 값 = 처음 보는 분야의 정직한 기대치.",
"do_not_cite": "self_auc (in-sample). 성능은 cv_auc 로만 말한다.",
"curve_probe_file": "ztc_curve_probe.npz"
}