Shar137/two / evaluation /SCORES_BF16.json
Shar137's picture
download
raw
4.28 kB
{
"tier": "BF16",
"weights": "Coder390 merged BF16 weights",
"protocol": "双卡 RTX PRO 6000,每卡 C8,上下文 100K,生成上限 94208,客户端不计时,DFLASH2,SGLang;GPQA198/MMLU500/LCB100 全量;空答记错、截断答对记对",
"public_scores": {
"BF16": "178/446/90",
"动态FP8": "178/445/90",
"静态FP8": "178/450/90"
},
"detail": {
"gpqa": {
"BF16": {
"n": 198,
"total": 198,
"correct": 178,
"p50": 3196,
"p70": 8200,
"p90": 28417,
"trunc94k": 2,
"empty": 2,
"buckets": {
"<2K": "80/85",
"2–12K": "66/68",
"12–24K": "17/20",
"24–48K": "12/18",
"≥48K": "3/7"
}
},
"动态FP8": {
"n": 198,
"total": 198,
"correct": 178,
"p50": 2815,
"p70": 8371,
"p90": 27744,
"trunc94k": 1,
"empty": 1,
"buckets": {
"<2K": "78/81",
"2–12K": "68/70",
"12–24K": "18/21",
"24–48K": "11/17",
"≥48K": "3/9"
}
},
"静态FP8": {
"n": 198,
"total": 198,
"correct": 178,
"p50": 2966,
"p70": 7832,
"p90": 26431,
"trunc94k": 1,
"empty": 2,
"buckets": {
"<2K": "82/84",
"2–12K": "68/74",
"12–24K": "17/19",
"24–48K": "9/17",
"≥48K": "2/4"
}
}
},
"mmlu": {
"BF16": {
"n": 500,
"total": 500,
"correct": 446,
"p50": 154,
"p70": 252,
"p90": 779,
"trunc94k": 1,
"empty": 0,
"buckets": {
"<2K": "432/472",
"2–12K": "13/23",
"12–24K": "1/3",
"24–48K": "0/0",
"≥48K": "0/2"
}
},
"动态FP8": {
"n": 500,
"total": 500,
"correct": 445,
"p50": 138,
"p70": 271,
"p90": 881,
"trunc94k": 1,
"empty": 0,
"buckets": {
"<2K": "436/477",
"2–12K": "8/21",
"12–24K": "0/1",
"24–48K": "0/0",
"≥48K": "1/1"
}
},
"静态FP8": {
"n": 500,
"total": 500,
"correct": 450,
"p50": 154,
"p70": 279,
"p90": 937,
"trunc94k": 1,
"empty": 0,
"buckets": {
"<2K": "435/469",
"2–12K": "13/27",
"12–24K": "2/3",
"24–48K": "0/0",
"≥48K": "0/1"
}
}
},
"lcb": {
"BF16": {
"n": 100,
"total": 100,
"correct": 90,
"p50": 5073,
"p70": 16033,
"p90": 38911,
"trunc94k": 2,
"empty": 2,
"buckets": {
"<2K": "40/40",
"2–12K": "25/26",
"12–24K": "11/14",
"24–48K": "9/11",
"≥48K": "5/9"
}
},
"动态FP8": {
"n": 100,
"total": 100,
"correct": 90,
"p50": 4791,
"p70": 15866,
"p90": 41168,
"trunc94k": 4,
"empty": 4,
"buckets": {
"<2K": "38/39",
"2–12K": "25/26",
"12–24K": "11/14",
"24–48K": "12/13",
"≥48K": "4/8"
}
},
"静态FP8": {
"n": 100,
"total": 100,
"correct": 90,
"p50": 4511,
"p70": 16027,
"p90": 43992,
"trunc94k": 3,
"empty": 3,
"buckets": {
"<2K": "39/39",
"2–12K": "24/26",
"12–24K": "13/13",
"24–48K": "12/14",
"≥48K": "2/8"
}
}
}
},
"evidence_dirs": {
"BF16": "eval-Coder390-bf16-pure-dual-c8-100k-notimeout-20261003",
"动态FP8": "eval-Coder390-dynfp8-dual-c8-100k-notimeout-20261002(GPQA、LCB)+ eval-Coder390-dynfp8-mmlu-retest2-20261003(MMLU)",
"静态FP8": "eval-Coder390-staticfp8-block128-dual-c8-100k-notimeout-20261004"
},
"note": "动态 FP8 是对本 BF16 权重在推理时做动态 FP8,不另出包;对外写 178/445/90。性能表与静态 FP8 共用 BENCH_STATICFP8"
}

Xet Storage Details

Size:
4.28 kB
·
Xet hash:
a519d7ef9babe833460f9f6f8eb805a981ff8b0754f3ecdfceb15765f9a4028a

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