| { | |
| "tier": "NVFP4_W4A16", | |
| "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 170/198，MMLU 439/500，LCB 91/100", | |
| "detail": { | |
| "gpqa": { | |
| "n": 198, | |
| "total": 198, | |
| "correct": 170, | |
| "p50": 2965, | |
| "p70": 9667, | |
| "p90": 28433, | |
| "trunc94k": 2, | |
| "empty": 3, | |
| "buckets": { | |
| "<2K": "77/80", | |
| "2–12K": "64/70", | |
| "12–24K": "17/21", | |
| "24–48K": "8/17", | |
| "≥48K": "4/10" | |
| } | |
| }, | |
| "mmlu": { | |
| "n": 500, | |
| "total": 500, | |
| "correct": 439, | |
| "p50": 159, | |
| "p70": 277, | |
| "p90": 1061, | |
| "trunc94k": 1, | |
| "empty": 0, | |
| "buckets": { | |
| "<2K": "427/471", | |
| "2–12K": "11/26", | |
| "12–24K": "0/1", | |
| "24–48K": "1/1", | |
| "≥48K": "0/1" | |
| } | |
| }, | |
| "lcb": { | |
| "n": 100, | |
| "total": 100, | |
| "correct": 91, | |
| "p50": 5119, | |
| "p70": 15790, | |
| "p90": 47768, | |
| "trunc94k": 4, | |
| "empty": 3, | |
| "buckets": { | |
| "<2K": "38/38", | |
| "2–12K": "26/27", | |
| "12–24K": "13/13", | |
| "24–48K": "10/12", | |
| "≥48K": "4/10" | |
| } | |
| } | |
| }, | |
| "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": true, | |
| "reasoning_tokens": 94209, | |
| "finish_reason": "length", | |
| "content": "pred 为空，生成顶到 94208" | |
| }, | |
| { | |
| "id": 79, | |
| "empty": true, | |
| "trunc94k": true, | |
| "reasoning_tokens": 94209, | |
| "finish_reason": "length", | |
| "content": "pred 为空，生成顶到 94208" | |
| }, | |
| { | |
| "id": 101, | |
| "empty": true, | |
| "trunc94k": false, | |
| "reasoning_tokens": 57565, | |
| "finish_reason": "stop", | |
| "content": "正常停，但 pred 为空" | |
| } | |
| ], | |
| "mmlu": [ | |
| { | |
| "id": "test:professional_law:784", | |
| "empty": false, | |
| "trunc94k": true, | |
| "ok": false, | |
| "reasoning_tokens": 94210, | |
| "finish_reason": "length", | |
| "content": "截断后仍抽出选项 B，判错，不是空答" | |
| } | |
| ], | |
| "lcb": [ | |
| { | |
| "id": 7, | |
| "empty": true, | |
| "trunc94k": true, | |
| "note": "empty_code", | |
| "reasoning_tokens": 94209 | |
| }, | |
| { | |
| "id": 6, | |
| "empty": true, | |
| "trunc94k": true, | |
| "note": "empty_code", | |
| "reasoning_tokens": 94209 | |
| }, | |
| { | |
| "id": 54, | |
| "empty": true, | |
| "trunc94k": true, | |
| "note": "empty_code", | |
| "reasoning_tokens": 94211 | |
| }, | |
| { | |
| "id": 53, | |
| "empty": false, | |
| "trunc94k": true, | |
| "note": "wrong_answer_case:0", | |
| "code_len": 20, | |
| "reasoning_tokens": 94194, | |
| "content": "有 20 字符代码，截断，不是空答" | |
| } | |
| ] | |
| } | |
| } | |
Xet Storage Details
- Size:
- 3.32 kB
- Xet hash:
- b18306d0a0fd476825ff305b26109776d92ba4586e0edc3a10fa23ea98219320
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.