reports/review/step2500.md
5.3 KB · 104 lines · markdown Raw
1 # Evaluation — reports/review/step2500_ckpt
2
3 base: `/root/models/Qwen3.5-9B` · 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 | 0.9917 | 0.8236 | ✅ |
12 | noul_brier_soft | <= 0.1 | 0.0028 | 0.0960 | ✅ |
13 | choice_top1 | >= 0.9 | 0.8646 | 0.5323 | ❌ |
14 | score_mae_expected | <= 0.35 | 0.1425 | 1.1297 | ✅ |
15 | kl | <= 0.15 | 0.0369 | 0.5100 | ✅ |
16 | ece | <= 0.03 | 0.0041 | 0.0942 | ✅ |
17 | ood KL ratio (ood/test_set_30k) | <= 2.0 | 6.847 | — | ❌ |
18
19 **excluding yuri_v1 exact-uniform placeholders (D1)**
20
21 | metric | target | value | B0 | pass |
22 |---|---|---|---|---|
23 | noul_auroc | >= 0.95 | 0.9917 | nan | ✅ |
24 | noul_brier_soft | <= 0.1 | 0.0034 | nan | ✅ |
25 | choice_top1 | >= 0.9 | 0.8646 | nan | ❌ |
26 | score_mae_expected | <= 0.35 | 0.1425 | nan | ✅ |
27 | kl | <= 0.15 | 0.0399 | nan | ✅ |
28 | ece | <= 0.03 | 0.0039 | nan | ✅ |
29 | ood KL ratio (ood/test_set_30k) | <= 2.0 | 6.335 | — | ❌ |
30
31 ## test_set_30k
32
33 all rows: `n=29955 · kl=0.0369 · js=0.0094 · top1=0.8863 · ece=0.0041 · mce=0.0089 · noul_auroc=0.9917 · noul_brier_soft=0.0028 · noul_brier_hard=0.0793 · score_mae_expected=0.1425 · score_rps=0.0162 · choice_top1=0.8646 · choice_kl=0.0705`
34
35 excluding yuri_v1 placeholders (2260 rows): `n=27695 · kl=0.0399 · js=0.0102 · top1=0.8863 · ece=0.0039 · mce=0.0089 · noul_auroc=0.9917 · noul_brier_soft=0.0034 · noul_brier_hard=0.0793 · score_mae_expected=0.1425 · score_rps=0.0162 · choice_top1=0.8646 · choice_kl=0.0705`
36
37 throughput: 25861 tok/s · 242.5 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 | 0.071 | 0.865 | 0.013 | — | — | — | 0.865 |
44 | noul | 12229 | 0.008 | 0.954 | 0.005 | 0.992 | 0.003 | — | — |
45 | score | 8527 | 0.042 | 0.832 | 0.008 | — | — | 0.143 | — |
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 | 0.248 | 0.854 | 0.027 | — | — | — | 0.854 |
52 | openjev_v2 | noul | 1432 | 0.022 | 0.994 | 0.004 | 1.000 | 0.005 | — | — |
53 | yuri_v1 | noul | 2260 | 0.000 | — | 0.010 | — | 0.000 | — | — |
54 | yuri_v3 | choice | 8312 | 0.052 | 0.866 | 0.014 | — | — | — | 0.866 |
55 | yuri_v3 | noul | 8537 | 0.008 | 0.947 | 0.005 | 0.990 | 0.003 | — | — |
56 | yuri_v3 | score | 8527 | 0.042 | 0.832 | 0.008 | — | — | 0.143 | — |
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 | 0.035 | 0.876 | 0.008 | 0.991 | 0.004 | 0.150 | 0.849 |
63 | biology | 1654 | 0.032 | 0.876 | 0.010 | 0.997 | 0.003 | 0.186 | 0.881 |
64 | business | 2352 | 0.031 | 0.908 | 0.006 | 0.995 | 0.002 | 0.104 | 0.863 |
65 | chemistry | 1333 | 0.028 | 0.919 | 0.011 | 0.995 | 0.003 | 0.131 | 0.939 |
66 | genomics | 1914 | 0.036 | 0.864 | 0.005 | 0.990 | 0.004 | 0.143 | 0.848 |
67 | knowledge | 4807 | 0.019 | 0.879 | 0.009 | 0.978 | 0.001 | 0.156 | 0.874 |
68 | medical | 2120 | 0.034 | 0.882 | 0.007 | 0.994 | 0.003 | 0.116 | 0.844 |
69 | openjev | 2319 | 0.108 | 0.941 | 0.013 | 1.000 | 0.005 | — | 0.854 |
70 | physics | 1656 | 0.031 | 0.862 | 0.008 | 0.990 | 0.002 | 0.137 | 0.869 |
71 | science | 1667 | 0.035 | 0.871 | 0.012 | 0.990 | 0.003 | 0.160 | 0.884 |
72 | spatial | 1647 | 0.033 | 0.878 | 0.009 | 0.982 | 0.003 | 0.155 | 0.840 |
73 | structured | 1668 | 0.035 | 0.864 | 0.006 | 0.981 | 0.005 | 0.180 | 0.843 |
74 | technical | 2581 | 0.034 | 0.901 | 0.008 | 0.992 | 0.003 | 0.110 | 0.853 |
75 | theology | 1654 | 0.033 | 0.871 | 0.005 | 0.990 | 0.003 | 0.153 | 0.901 |
76
77 ## ood
78
79 all rows: `n=13058 · kl=0.2527 · js=0.0443 · top1=0.9224 · ece=0.0393 · mce=0.1853 · noul_auroc=0.9843 · noul_brier_soft=0.0428 · noul_brier_hard=0.0428 · score_mae_expected=0.8291 · score_rps=0.5762 · choice_top1=0.8537 · choice_kl=0.3757`
80
81 excluding yuri_v1 placeholders (0 rows): `n=13058 · kl=0.2527 · js=0.0443 · top1=0.9224 · ece=0.0393 · mce=0.1853 · noul_auroc=0.9843 · noul_brier_soft=0.0428 · noul_brier_hard=0.0428 · score_mae_expected=0.8291 · score_rps=0.5762 · choice_top1=0.8537 · choice_kl=0.3757`
82
83 throughput: 27227 tok/s · 102.7 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 | 0.376 | 0.854 | 0.065 | — | — | — | 0.854 |
90 | noul | 9767 | 0.204 | 0.949 | 0.030 | 0.984 | 0.043 | — | — |
91 | score | 72 | 1.367 | 0.431 | 0.207 | — | — | 0.829 | — |
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 | 0.376 | 0.854 | 0.065 | — | — | — | 0.854 |
98 | openjev_v2 | noul | 9767 | 0.204 | 0.949 | 0.030 | 0.984 | 0.043 | — | — |
99 | openjev_v2 | score | 72 | 1.367 | 0.431 | 0.207 | — | — | 0.829 | — |
100
101 ## Choice permutation consistency
102
103 `{"n_rows": 500, "mean_max_abs_diff": 0.03688684105873108, "p90_max_abs_diff": 0.07897255569696426, "top1_flip_rate": 0.053}`
104