Shar137/two / evaluation /SCORES_NVFP4_MIXED.json
Shar137's picture
download
raw
3.06 kB
{
"tier": "NVFP4_MIXED",
"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 172/198,MMLU 442/500,LCB 88/100",
"detail": {
"gpqa": {
"n": 198,
"total": 198,
"correct": 172,
"p50": 2953,
"p70": 10010,
"p90": 35731,
"trunc94k": 2,
"empty": 4,
"buckets": {
"<2K": "79/82",
"2–12K": "58/63",
"12–24K": "19/22",
"24–48K": "14/22",
"≥48K": "2/9"
}
},
"mmlu": {
"n": 500,
"total": 500,
"correct": 442,
"p50": 157,
"p70": 273,
"p90": 778,
"trunc94k": 0,
"empty": 0,
"buckets": {
"<2K": "432/477",
"2–12K": "8/20",
"12–24K": "1/2",
"24–48K": "1/1",
"≥48K": "0/0"
}
},
"lcb": {
"n": 100,
"total": 100,
"correct": 88,
"p50": 4731,
"p70": 15122,
"p90": 41344,
"trunc94k": 2,
"empty": 1,
"buckets": {
"<2K": "40/41",
"2–12K": "21/24",
"12–24K": "16/16",
"24–48K": "7/10",
"≥48K": "4/9"
}
}
},
"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。分桶按 reasoning_tokens 左闭右开(2000/12000/24000/48000)。",
"stats_source": "与静态 FP8 同一口径的同题对比;分桶与异常题由 results.jsonl 按同一定义复算,P50/P70/P90、94K 截断、空答与同题对比逐项一致",
"stats_updated_at": "2026-10-05T12:35:40+08:00",
"anomalies": {
"gpqa": [
{
"id": 24,
"empty": true,
"trunc94k": false,
"reasoning_tokens": 60716,
"finish_reason": "stop",
"content": "正常停,pred 为空"
},
{
"id": 74,
"empty": true,
"trunc94k": true,
"reasoning_tokens": 94210,
"finish_reason": "length",
"content": "pred 为空,生成顶到 94208"
},
{
"id": 79,
"empty": true,
"trunc94k": true,
"reasoning_tokens": 94208,
"finish_reason": "length",
"content": "pred 为空,生成顶到 94208"
},
{
"id": 127,
"empty": true,
"trunc94k": false,
"reasoning_tokens": 25676,
"finish_reason": "stop",
"content": "正常停,pred 为空"
}
],
"mmlu": [],
"lcb": [
{
"id": 11,
"empty": true,
"trunc94k": true,
"note": "empty_code",
"reasoning_tokens": 94209
},
{
"id": 6,
"empty": false,
"trunc94k": true,
"note": "截断,有代码(2338 字符),答错",
"reasoning_tokens": 93343
}
]
}
}

Xet Storage Details

Size:
3.06 kB
·
Xet hash:
f79bab51e1d53edc7a188325bd10d78a62358452dc781cb954fd58302c62c17b

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.