| { | |
| "tier": "NVFP4_W4A4", | |
| "protocol": "单卡 RTX PRO 6000，C8，上下文 100K，客户端不计时，DFLASH2，SGLang；GPQA198/MMLU500/LCB100 全量", | |
| "public_scores": "GPQA 167/198，MMLU 440/500，LCB 92/100", | |
| "detail": { | |
| "gpqa": { | |
| "n": 198, | |
| "total": 198, | |
| "correct": 167, | |
| "p50": 2876, | |
| "p70": 10996, | |
| "p90": 31986, | |
| "trunc94k": 2, | |
| "empty": 3, | |
| "buckets": { | |
| "<2K": "76/82", | |
| "2–12K": "58/60", | |
| "12–24K": "20/23", | |
| "24–48K": "12/26", | |
| "≥48K": "1/7" | |
| } | |
| }, | |
| "mmlu": { | |
| "n": 500, | |
| "total": 500, | |
| "correct": 440, | |
| "p50": 168, | |
| "p70": 306, | |
| "p90": 967, | |
| "trunc94k": 2, | |
| "empty": 0, | |
| "buckets": { | |
| "<2K": "426/469", | |
| "2–12K": "13/28", | |
| "12–24K": "0/1", | |
| "24–48K": "0/0", | |
| "≥48K": "1/2" | |
| } | |
| }, | |
| "lcb": { | |
| "n": 100, | |
| "total": 100, | |
| "correct": 92, | |
| "p50": 6338, | |
| "p70": 15771, | |
| "p90": 54096, | |
| "trunc94k": 3, | |
| "empty": 3, | |
| "buckets": { | |
| "<2K": "33/33", | |
| "2–12K": "30/31", | |
| "12–24K": "9/11", | |
| "24–48K": "13/14", | |
| "≥48K": "7/11" | |
| } | |
| } | |
| }, | |
| "definition": "思考长度取 usage.reasoning_tokens。P50/P70/P90 用 verify_table.py 的线性插值。94K 截断是 finish_reason=length 或 completion_tokens>=94208。空答：GPQA 没有 pred，MMLU 没有 prediction，LCB 没有 code_len。截断后仍抽出选项的不算空答。", | |
| "stats_source": "verify_table.py，与静态 FP8 同一函数", | |
| "stats_updated_at": "2026-10-05T10:15:03+08:00", | |
| "anomalies": { | |
| "gpqa": [ | |
| { | |
| "id": 10, | |
| "empty": true, | |
| "trunc94k": false, | |
| "reasoning_tokens": 28813, | |
| "finish_reason": "stop", | |
| "content": "正常停，pred 为空" | |
| }, | |
| { | |
| "id": 79, | |
| "empty": true, | |
| "trunc94k": true, | |
| "reasoning_tokens": 94213, | |
| "finish_reason": "length", | |
| "content": "pred 为空，生成顶到 94208" | |
| }, | |
| { | |
| "id": 145, | |
| "empty": true, | |
| "trunc94k": true, | |
| "reasoning_tokens": 94211, | |
| "finish_reason": "length", | |
| "content": "pred 为空，生成顶到 94208" | |
| } | |
| ], | |
| "mmlu": [ | |
| { | |
| "id": "test:clinical_knowledge:140", | |
| "empty": false, | |
| "trunc94k": true, | |
| "ok": true, | |
| "prediction": "D", | |
| "content": "截断但抽出 D，判对" | |
| }, | |
| { | |
| "id": "test:electrical_engineering:137", | |
| "empty": false, | |
| "trunc94k": true, | |
| "ok": false, | |
| "prediction": "C", | |
| "content": "截断后抽出 C，判错" | |
| } | |
| ], | |
| "lcb": [ | |
| { | |
| "id": 0, | |
| "empty": true, | |
| "trunc94k": true, | |
| "note": "empty_code", | |
| "reasoning_tokens": 94214 | |
| }, | |
| { | |
| "id": 53, | |
| "empty": true, | |
| "trunc94k": true, | |
| "note": "empty_code", | |
| "reasoning_tokens": 94212 | |
| }, | |
| { | |
| "id": 96, | |
| "empty": true, | |
| "trunc94k": true, | |
| "note": "empty_code", | |
| "reasoning_tokens": 94210 | |
| } | |
| ] | |
| } | |
| } | |
Xet Storage Details
- Size:
- 3.21 kB
- Xet hash:
- f8b674817a8f083b225bf5891b5cf2783a71b663f7acf545a212d86b6a3525fd
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.