reports/review/step0500_27b.md
5.3 KB · 104 lines · markdown Raw
1 # Evaluation — reports/review/step0500_27b_ckpt
2
3 base: `/root/models/Qwen3.8-27B` · temperature: yes · max_seq_len 1024
4
5 ## Acceptance (DESIGN §8.3, main = test_set_30k)
6
7 **all rows**
8
9 | metric | target | value | B0 | pass |
10 |---|---|---|---|---|
11 | noul_auroc | >= 0.95 | 0.9838 | 0.8763 | ✅ |
12 | noul_brier_soft | <= 0.1 | 0.0063 | 0.0628 | ✅ |
13 | choice_top1 | >= 0.9 | 0.8278 | 0.5814 | ❌ |
14 | score_mae_expected | <= 0.35 | 0.1987 | 0.8420 | ✅ |
15 | kl | <= 0.15 | 0.0579 | 0.4296 | ✅ |
16 | ece | <= 0.03 | 0.0032 | 0.0501 | ✅ |
17 | ood KL ratio (ood/test_set_30k) | <= 2.0 | 2.987 | — | ❌ |
18
19 **excluding yuri_v1 exact-uniform placeholders (D1)**
20
21 | metric | target | value | B0 | pass |
22 |---|---|---|---|---|
23 | noul_auroc | >= 0.95 | 0.9838 | 0.8763 | ✅ |
24 | noul_brier_soft | <= 0.1 | 0.0077 | 0.0552 | ✅ |
25 | choice_top1 | >= 0.9 | 0.8278 | 0.5814 | ❌ |
26 | score_mae_expected | <= 0.35 | 0.1987 | 0.8420 | ✅ |
27 | kl | <= 0.15 | 0.0626 | 0.4411 | ✅ |
28 | ece | <= 0.03 | 0.0028 | 0.0307 | ✅ |
29 | ood KL ratio (ood/test_set_30k) | <= 2.0 | 2.763 | — | ❌ |
30
31 ## test_set_30k
32
33 all rows: `n=29955 · kl=0.0579 · js=0.0147 · top1=0.8576 · ece=0.0032 · mce=0.0186 · noul_auroc=0.9838 · noul_brier_soft=0.0063 · noul_brier_hard=0.0877 · score_mae_expected=0.1987 · score_rps=0.0280 · choice_top1=0.8278 · choice_kl=0.1011`
34
35 excluding yuri_v1 placeholders (2260 rows): `n=27695 · kl=0.0626 · js=0.0159 · top1=0.8576 · ece=0.0028 · mce=0.0186 · noul_auroc=0.9838 · noul_brier_soft=0.0077 · noul_brier_hard=0.0877 · score_mae_expected=0.1987 · score_rps=0.0280 · choice_top1=0.8278 · choice_kl=0.1011`
36
37 throughput: 9119 tok/s · 85.5 rows/s
38
39 ### test_set_30k by kind
40
41 | kind | n | kl | top1 | ece | noul_auroc | noul_brier_soft | score_mae_expected | choice_top1 |
42 |---|---|---|---|---|---|---|---|---|
43 | choice | 9199 | 0.101 | 0.828 | 0.011 | — | — | — | 0.828 |
44 | noul | 12229 | 0.019 | 0.935 | 0.008 | 0.984 | 0.006 | — | — |
45 | score | 8527 | 0.067 | 0.800 | 0.007 | — | — | 0.199 | — |
46
47 ### test_set_30k by source × kind
48
49 | source | kind | n | kl | top1 | ece | noul_auroc | noul_brier_soft | score_mae_expected | choice_top1 |
50 |---|---|---|---|---|---|---|---|---|---|
51 | openjev_v2 | choice | 887 | 0.369 | 0.788 | 0.045 | — | — | — | 0.788 |
52 | openjev_v2 | noul | 1432 | 0.084 | 0.969 | 0.020 | 0.996 | 0.023 | — | — |
53 | yuri_v1 | noul | 2260 | 0.000 | — | 0.009 | — | 0.000 | — | — |
54 | yuri_v3 | choice | 8312 | 0.072 | 0.832 | 0.008 | — | — | — | 0.832 |
55 | yuri_v3 | noul | 8537 | 0.013 | 0.930 | 0.011 | 0.985 | 0.005 | — | — |
56 | yuri_v3 | score | 8527 | 0.067 | 0.800 | 0.007 | — | — | 0.199 | — |
57
58 ### test_set_30k by family
59
60 | family | n | kl | top1 | ece | noul_auroc | noul_brier_soft | score_mae_expected | choice_top1 |
61 |---|---|---|---|---|---|---|---|---|
62 | agent | 2583 | 0.048 | 0.845 | 0.007 | 0.992 | 0.004 | 0.224 | 0.803 |
63 | biology | 1654 | 0.046 | 0.855 | 0.008 | 0.992 | 0.006 | 0.208 | 0.848 |
64 | business | 2352 | 0.048 | 0.882 | 0.009 | 0.991 | 0.003 | 0.150 | 0.839 |
65 | chemistry | 1333 | 0.046 | 0.887 | 0.009 | 0.992 | 0.005 | 0.194 | 0.904 |
66 | genomics | 1914 | 0.061 | 0.830 | 0.011 | 0.978 | 0.007 | 0.210 | 0.817 |
67 | knowledge | 4807 | 0.027 | 0.848 | 0.007 | 0.966 | 0.002 | 0.204 | 0.846 |
68 | medical | 2120 | 0.051 | 0.866 | 0.008 | 0.989 | 0.005 | 0.154 | 0.789 |
69 | openjev | 2319 | 0.193 | 0.901 | 0.026 | 0.996 | 0.023 | — | 0.788 |
70 | physics | 1656 | 0.049 | 0.837 | 0.008 | 0.984 | 0.004 | 0.190 | 0.858 |
71 | science | 1667 | 0.051 | 0.841 | 0.010 | 0.980 | 0.004 | 0.222 | 0.842 |
72 | spatial | 1647 | 0.052 | 0.844 | 0.010 | 0.986 | 0.004 | 0.217 | 0.791 |
73 | structured | 1668 | 0.049 | 0.839 | 0.010 | 0.972 | 0.008 | 0.249 | 0.828 |
74 | technical | 2581 | 0.054 | 0.872 | 0.008 | 0.991 | 0.005 | 0.158 | 0.824 |
75 | theology | 1654 | 0.050 | 0.845 | 0.009 | 0.988 | 0.005 | 0.238 | 0.866 |
76
77 ## ood
78
79 all rows: `n=13058 · kl=0.1730 · js=0.0434 · top1=0.9283 · ece=0.0078 · mce=0.2500 · noul_auroc=0.9869 · noul_brier_soft=0.0336 · noul_brier_hard=0.0336 · score_mae_expected=0.6120 · score_rps=0.4686 · choice_top1=0.8472 · choice_kl=0.3079`
80
81 excluding yuri_v1 placeholders (0 rows): `n=13058 · kl=0.1730 · js=0.0434 · top1=0.9283 · ece=0.0078 · mce=0.2500 · noul_auroc=0.9869 · noul_brier_soft=0.0336 · noul_brier_hard=0.0336 · score_mae_expected=0.6120 · score_rps=0.4686 · choice_top1=0.8472 · choice_kl=0.3079`
82
83 throughput: 9422 tok/s · 35.5 rows/s
84
85 ### ood by kind
86
87 | kind | n | kl | top1 | ece | noul_auroc | noul_brier_soft | score_mae_expected | choice_top1 |
88 |---|---|---|---|---|---|---|---|---|
89 | choice | 3219 | 0.308 | 0.847 | 0.033 | — | — | — | 0.847 |
90 | noul | 9767 | 0.122 | 0.958 | 0.013 | 0.987 | 0.034 | — | — |
91 | score | 72 | 1.105 | 0.583 | 0.176 | — | — | 0.612 | — |
92
93 ### ood by source × kind
94
95 | source | kind | n | kl | top1 | ece | noul_auroc | noul_brier_soft | score_mae_expected | choice_top1 |
96 |---|---|---|---|---|---|---|---|---|---|
97 | openjev_v2 | choice | 3219 | 0.308 | 0.847 | 0.033 | — | — | — | 0.847 |
98 | openjev_v2 | noul | 9767 | 0.122 | 0.958 | 0.013 | 0.987 | 0.034 | — | — |
99 | openjev_v2 | score | 72 | 1.105 | 0.583 | 0.176 | — | — | 0.612 | — |
100
101 ## Choice permutation consistency
102
103 `{"n_rows": 500, "mean_max_abs_diff": 0.06085456162691116, "p90_max_abs_diff": 0.13134749233722687, "top1_flip_rate": 0.086}`
104