reports/review/step0500.md
5.3 KB · 104 lines · markdown Raw
1 # Evaluation — reports/review/step0500_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 | 0.9772 | 0.8236 | ✅ |
12 | noul_brier_soft | <= 0.1 | 0.0174 | 0.0960 | ✅ |
13 | choice_top1 | >= 0.9 | 0.8168 | 0.5323 | ❌ |
14 | score_mae_expected | <= 0.35 | 0.2244 | 1.1297 | ✅ |
15 | kl | <= 0.15 | 0.0809 | 0.5100 | ✅ |
16 | ece | <= 0.03 | 0.0215 | 0.0942 | ✅ |
17 | ood KL ratio (ood/test_set_30k) | <= 2.0 | 3.658 | — | ❌ |
18
19 **excluding yuri_v1 exact-uniform placeholders (D1)**
20
21 | metric | target | value | B0 | pass |
22 |---|---|---|---|---|
23 | noul_auroc | >= 0.95 | 0.9772 | nan | ✅ |
24 | noul_brier_soft | <= 0.1 | 0.0126 | nan | ✅ |
25 | choice_top1 | >= 0.9 | 0.8168 | nan | ❌ |
26 | score_mae_expected | <= 0.35 | 0.2244 | nan | ✅ |
27 | kl | <= 0.15 | 0.0806 | nan | ✅ |
28 | ece | <= 0.03 | 0.0074 | nan | ✅ |
29 | ood KL ratio (ood/test_set_30k) | <= 2.0 | 3.670 | — | ❌ |
30
31 ## test_set_30k
32
33 all rows: `n=29955 · kl=0.0809 · js=0.0207 · top1=0.8240 · ece=0.0215 · mce=0.0762 · noul_auroc=0.9772 · noul_brier_soft=0.0174 · noul_brier_hard=0.0927 · score_mae_expected=0.2244 · score_rps=0.0369 · choice_top1=0.8168 · choice_kl=0.1185`
34
35 excluding yuri_v1 placeholders (2260 rows): `n=27695 · kl=0.0806 · js=0.0208 · top1=0.8240 · ece=0.0074 · mce=0.0526 · noul_auroc=0.9772 · noul_brier_soft=0.0126 · noul_brier_hard=0.0927 · score_mae_expected=0.2244 · score_rps=0.0369 · choice_top1=0.8168 · choice_kl=0.1185`
36
37 throughput: 25871 tok/s · 242.6 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 | 0.118 | 0.817 | 0.004 | — | — | — | 0.817 |
44 | noul | 12229 | 0.046 | 0.903 | 0.042 | 0.977 | 0.017 | — | — |
45 | score | 8527 | 0.090 | 0.741 | 0.016 | — | — | 0.224 | — |
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 | 0.405 | 0.758 | 0.044 | — | — | — | 0.758 |
52 | openjev_v2 | noul | 1432 | 0.128 | 0.964 | 0.024 | 0.986 | 0.034 | — | — |
53 | yuri_v1 | noul | 2260 | 0.084 | — | 0.193 | — | 0.038 | — | — |
54 | yuri_v3 | choice | 8312 | 0.088 | 0.823 | 0.004 | — | — | — | 0.823 |
55 | yuri_v3 | noul | 8537 | 0.022 | 0.892 | 0.008 | 0.978 | 0.009 | — | — |
56 | yuri_v3 | score | 8527 | 0.090 | 0.741 | 0.016 | — | — | 0.224 | — |
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 | 0.064 | 0.826 | 0.008 | 0.981 | 0.009 | 0.244 | 0.827 |
63 | biology | 1654 | 0.071 | 0.809 | 0.016 | 0.990 | 0.011 | 0.277 | 0.822 |
64 | business | 2352 | 0.062 | 0.855 | 0.009 | 0.987 | 0.007 | 0.168 | 0.854 |
65 | chemistry | 1333 | 0.056 | 0.873 | 0.017 | 0.990 | 0.006 | 0.223 | 0.915 |
66 | genomics | 1914 | 0.076 | 0.804 | 0.007 | 0.978 | 0.008 | 0.257 | 0.796 |
67 | knowledge | 4807 | 0.077 | 0.811 | 0.102 | 0.953 | 0.031 | 0.213 | 0.803 |
68 | medical | 2120 | 0.069 | 0.823 | 0.016 | 0.984 | 0.008 | 0.179 | 0.797 |
69 | openjev | 2319 | 0.234 | 0.886 | 0.020 | 0.986 | 0.034 | — | 0.758 |
70 | physics | 1656 | 0.062 | 0.802 | 0.009 | 0.976 | 0.009 | 0.214 | 0.841 |
71 | science | 1667 | 0.060 | 0.805 | 0.023 | 0.974 | 0.009 | 0.215 | 0.844 |
72 | spatial | 1647 | 0.069 | 0.822 | 0.012 | 0.972 | 0.007 | 0.245 | 0.782 |
73 | structured | 1668 | 0.068 | 0.785 | 0.015 | 0.961 | 0.013 | 0.278 | 0.770 |
74 | technical | 2581 | 0.069 | 0.818 | 0.009 | 0.985 | 0.009 | 0.186 | 0.808 |
75 | theology | 1654 | 0.065 | 0.804 | 0.010 | 0.983 | 0.009 | 0.265 | 0.874 |
76
77 ## ood
78
79 all rows: `n=13058 · kl=0.2958 · js=0.0782 · top1=0.8789 · ece=0.0053 · mce=0.2488 · noul_auroc=0.9489 · noul_brier_soft=0.0647 · noul_brier_hard=0.0647 · score_mae_expected=0.6340 · score_rps=0.4477 · choice_top1=0.7645 · choice_kl=0.4942`
80
81 excluding yuri_v1 placeholders (0 rows): `n=13058 · kl=0.2958 · js=0.0782 · top1=0.8789 · ece=0.0053 · mce=0.2488 · noul_auroc=0.9489 · noul_brier_soft=0.0647 · noul_brier_hard=0.0647 · score_mae_expected=0.6340 · score_rps=0.4477 · choice_top1=0.7645 · choice_kl=0.4942`
82
83 throughput: 27224 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 | 0.494 | 0.765 | 0.047 | — | — | — | 0.765 |
90 | noul | 9767 | 0.224 | 0.919 | 0.013 | 0.949 | 0.065 | — | — |
91 | score | 72 | 1.151 | 0.611 | 0.179 | — | — | 0.634 | — |
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 | 0.494 | 0.765 | 0.047 | — | — | — | 0.765 |
98 | openjev_v2 | noul | 9767 | 0.224 | 0.919 | 0.013 | 0.949 | 0.065 | — | — |
99 | openjev_v2 | score | 72 | 1.151 | 0.611 | 0.179 | — | — | 0.634 | — |
100
101 ## Choice permutation consistency
102
103 `{"n_rows": 500, "mean_max_abs_diff": 0.06840353459119797, "p90_max_abs_diff": 0.14390206336975098, "top1_flip_rate": 0.1115}`
104