Text Classification
Transformers
Safetensors
English
Korean
qwen3_5
image-text-to-text
ztc
answer-verification
hallucination-detection
zero-token
confidence-estimation
Instructions to use FINAL-Bench/ZTC-Judge-27B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use FINAL-Bench/ZTC-Judge-27B with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-classification", model="FINAL-Bench/ZTC-Judge-27B")# pip install -U transformers accelerate # Load model directly from transformers import AutoProcessor, AutoModelForMultimodalLM processor = AutoProcessor.from_pretrained("FINAL-Bench/ZTC-Judge-27B") model = AutoModelForMultimodalLM.from_pretrained("FINAL-Bench/ZTC-Judge-27B", device_map="auto") - Notebooks
- Google Colab
- Kaggle
config: corrected per-domain LODO (curved probe), leaderboard 0.7282, baseline 0.6223
7491b79 verified Download ztc_config.json from FINAL-Bench/ZTC-Judge-27B: direct link, hf CLI and curl.
- Browser
- Download file 2.92 kB
-
https://huggingface.co/FINAL-Bench/ZTC-Judge-27B/resolve/main/ztc_config.json
- Command line
-
hf download hf://FINAL-Bench/ZTC-Judge-27B/ztc_config.json
-
curl -L -o ztc_config.json https://huggingface.co/FINAL-Bench/ZTC-Judge-27B/resolve/main/ztc_config.json
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" | |
| } |