JEV / code /reports /eval_v08_bundle.md
cloudyu's picture
v0.8.0 dual-head bundle: pristine Qwen3.5-9B backbone + lm_head (bit-identical, HumanEval 70.7% = base) + unmerged LoRA decision adapter + 24-slot head; card: one backbone two heads
448ef61 verified
|
Raw
History Blame Contribute Delete
3.77 kB

Evaluation — export exports/jev-judge-qwen35-9b-v0.8

base: /root/models/Qwen3.5-9B · temperature: yes · max_seq_len 1024

Acceptance (DESIGN §8.3, main = test_set_30k)

all rows

metric target value B0 pass
noul_auroc >= 0.95 0.9958 nan ✅
noul_brier_soft <= 0.1 0.0015 nan ✅
choice_top1 >= 0.9 0.8975 nan ❌
score_mae_expected <= 0.35 0.1030 nan ✅
kl <= 0.15 0.0211 nan ✅
ece <= 0.03 0.0008 nan ✅

excluding yuri_v1 exact-uniform placeholders (D1)

metric target value B0 pass
noul_auroc >= 0.95 0.9958 nan ✅
noul_brier_soft <= 0.1 0.0019 nan ✅
choice_top1 >= 0.9 0.8975 nan ❌
score_mae_expected <= 0.35 0.1030 nan ✅
kl <= 0.15 0.0228 nan ✅
ece <= 0.03 0.0010 nan ✅

test_set_30k

all rows: n=29955 · kl=0.0211 · js=0.0054 · top1=0.9176 · ece=0.0008 · mce=0.0557 · noul_auroc=0.9958 · noul_brier_soft=0.0015 · noul_brier_hard=0.0744 · score_mae_expected=0.1030 · score_rps=0.0084 · choice_top1=0.8975 · choice_kl=0.0423

excluding yuri_v1 placeholders (2260 rows): n=27695 · kl=0.0228 · js=0.0059 · top1=0.9176 · ece=0.0010 · mce=0.0557 · noul_auroc=0.9958 · noul_brier_soft=0.0019 · noul_brier_hard=0.0744 · score_mae_expected=0.1030 · score_rps=0.0084 · choice_top1=0.8975 · choice_kl=0.0423

throughput: 43375 tok/s · 406.7 rows/s

test_set_30k by kind

kind n kl top1 ece noul_auroc noul_brier_soft score_mae_expected choice_top1
choice 9199 0.042 0.897 0.002 — — — 0.897
noul 12229 0.004 0.967 0.002 0.996 0.002 — —
score 8527 0.023 0.882 0.002 — — 0.103 —

test_set_30k by source × kind

source kind n kl top1 ece noul_auroc noul_brier_soft score_mae_expected choice_top1
openjev_v2 choice 887 0.176 0.858 0.021 — — — 0.858
openjev_v2 noul 1432 0.004 0.997 0.003 1.000 0.001 — —
yuri_v1 noul 2260 0.000 — 0.005 — 0.000 — —
yuri_v3 choice 8312 0.028 0.902 0.002 — — — 0.902
yuri_v3 noul 8537 0.005 0.962 0.001 0.994 0.002 — —
yuri_v3 score 8527 0.023 0.882 0.002 — — 0.103 —

test_set_30k by family

family n kl top1 ece noul_auroc noul_brier_soft score_mae_expected choice_top1
agent 2583 0.018 0.915 0.003 0.996 0.002 0.104 0.897
biology 1654 0.018 0.912 0.006 0.999 0.002 0.130 0.922
business 2352 0.017 0.944 0.002 0.998 0.001 0.074 0.920
chemistry 1333 0.015 0.932 0.003 0.997 0.002 0.099 0.939
genomics 1914 0.019 0.907 0.003 0.995 0.002 0.104 0.892
knowledge 4807 0.011 0.908 0.004 0.992 0.001 0.116 0.893
medical 2120 0.018 0.922 0.005 0.996 0.002 0.080 0.891
openjev 2319 0.070 0.945 0.008 1.000 0.001 — 0.858
physics 1656 0.018 0.896 0.002 0.994 0.002 0.101 0.875
science 1667 0.019 0.902 0.006 0.992 0.002 0.118 0.913
spatial 1647 0.019 0.903 0.005 0.989 0.002 0.102 0.874
structured 1668 0.020 0.907 0.005 0.988 0.003 0.135 0.899
technical 2581 0.019 0.925 0.003 0.994 0.002 0.077 0.896
theology 1654 0.018 0.913 0.003 0.995 0.002 0.124 0.926