reports/eval_s2_9b.md
8.4 KB · 158 lines · markdown Raw
1 # Evaluation — checkpoints/s2_9b_cooldown/best
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 | 0.9938 | 0.8236 | ✅ |
12 | noul_brier_soft | <= 0.1 | 0.0022 | 0.0960 | ✅ |
13 | choice_top1 | >= 0.9 | 0.8836 | 0.5323 | ❌ |
14 | score_mae_expected | <= 0.35 | 0.1187 | 1.1297 | ✅ |
15 | kl | <= 0.15 | 0.0276 | 0.5100 | ✅ |
16 | ece | <= 0.03 | 0.0014 | 0.0942 | ✅ |
17 | ood KL ratio (ood/test_set_30k) | <= 2.0 | 7.526 | — | ❌ |
18
19 **excluding yuri_v1 exact-uniform placeholders (D1)**
20
21 | metric | target | value | B0 | pass |
22 |---|---|---|---|---|
23 | noul_auroc | >= 0.95 | 0.9938 | nan | ✅ |
24 | noul_brier_soft | <= 0.1 | 0.0026 | nan | ✅ |
25 | choice_top1 | >= 0.9 | 0.8836 | nan | ❌ |
26 | score_mae_expected | <= 0.35 | 0.1187 | nan | ✅ |
27 | kl | <= 0.15 | 0.0299 | nan | ✅ |
28 | ece | <= 0.03 | 0.0016 | nan | ✅ |
29 | ood KL ratio (ood/test_set_30k) | <= 2.0 | 6.966 | — | ❌ |
30
31 ## test_set_30k
32
33 all rows: `n=29955 · kl=0.0276 · js=0.0072 · top1=0.9036 · ece=0.0014 · mce=0.0079 · noul_auroc=0.9938 · noul_brier_soft=0.0022 · noul_brier_hard=0.0761 · score_mae_expected=0.1187 · score_rps=0.0113 · choice_top1=0.8836 · choice_kl=0.0549`
34
35 excluding yuri_v1 placeholders (2260 rows): `n=27695 · kl=0.0299 · js=0.0078 · top1=0.9036 · ece=0.0016 · mce=0.0079 · noul_auroc=0.9938 · noul_brier_soft=0.0026 · noul_brier_hard=0.0761 · score_mae_expected=0.1187 · score_rps=0.0113 · choice_top1=0.8836 · choice_kl=0.0549`
36
37 throughput: 25870 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 | 0.055 | 0.884 | 0.003 | — | — | — | 0.884 |
44 | noul | 12229 | 0.006 | 0.959 | 0.002 | 0.994 | 0.002 | — | — |
45 | score | 8527 | 0.030 | 0.861 | 0.002 | — | — | 0.119 | — |
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 | 0.220 | 0.858 | 0.018 | — | — | — | 0.858 |
52 | openjev_v2 | noul | 1432 | 0.010 | 0.997 | 0.004 | 1.000 | 0.003 | — | — |
53 | yuri_v1 | noul | 2260 | 0.000 | — | 0.009 | — | 0.000 | — | — |
54 | yuri_v3 | choice | 8312 | 0.037 | 0.886 | 0.002 | — | — | — | 0.886 |
55 | yuri_v3 | noul | 8537 | 0.006 | 0.953 | 0.001 | 0.992 | 0.003 | — | — |
56 | yuri_v3 | score | 8527 | 0.030 | 0.861 | 0.002 | — | — | 0.119 | — |
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 | 0.023 | 0.897 | 0.003 | 0.994 | 0.002 | 0.123 | 0.877 |
63 | biology | 1654 | 0.024 | 0.898 | 0.004 | 0.998 | 0.002 | 0.148 | 0.902 |
64 | business | 2352 | 0.022 | 0.930 | 0.003 | 0.997 | 0.002 | 0.089 | 0.904 |
65 | chemistry | 1333 | 0.020 | 0.931 | 0.005 | 0.997 | 0.002 | 0.113 | 0.946 |
66 | genomics | 1914 | 0.026 | 0.893 | 0.005 | 0.991 | 0.003 | 0.114 | 0.882 |
67 | knowledge | 4807 | 0.014 | 0.888 | 0.005 | 0.983 | 0.001 | 0.132 | 0.879 |
68 | medical | 2120 | 0.024 | 0.901 | 0.004 | 0.994 | 0.002 | 0.092 | 0.865 |
69 | openjev | 2319 | 0.090 | 0.945 | 0.007 | 1.000 | 0.003 | — | 0.858 |
70 | physics | 1656 | 0.023 | 0.881 | 0.004 | 0.992 | 0.002 | 0.117 | 0.869 |
71 | science | 1667 | 0.026 | 0.890 | 0.005 | 0.989 | 0.002 | 0.136 | 0.906 |
72 | spatial | 1647 | 0.025 | 0.886 | 0.007 | 0.984 | 0.002 | 0.119 | 0.850 |
73 | structured | 1668 | 0.025 | 0.885 | 0.004 | 0.983 | 0.004 | 0.152 | 0.881 |
74 | technical | 2581 | 0.025 | 0.918 | 0.004 | 0.993 | 0.002 | 0.089 | 0.878 |
75 | theology | 1654 | 0.025 | 0.893 | 0.005 | 0.992 | 0.002 | 0.144 | 0.909 |
76
77 ## test
78
79 all rows: `n=14261 · kl=0.0252 · js=0.0065 · top1=0.9076 · ece=0.0029 · mce=0.0243 · noul_auroc=0.9948 · noul_brier_soft=0.0018 · noul_brier_hard=0.0653 · score_mae_expected=0.1165 · score_rps=0.0111 · choice_top1=0.8751 · choice_kl=0.0622`
80
81 excluding yuri_v1 placeholders (2951 rows): `n=11310 · kl=0.0317 · js=0.0082 · top1=0.9076 · ece=0.0014 · mce=0.0243 · noul_auroc=0.9948 · noul_brier_soft=0.0029 · noul_brier_hard=0.0653 · score_mae_expected=0.1165 · score_rps=0.0111 · choice_top1=0.8751 · choice_kl=0.0622`
82
83 throughput: 26259 tok/s · 204.4 rows/s
84
85 ### test by kind
86
87 | kind | n | kl | top1 | ece | noul_auroc | noul_brier_soft | score_mae_expected | choice_top1 |
88 |---|---|---|---|---|---|---|---|---|
89 | choice | 3620 | 0.062 | 0.875 | 0.004 | — | — | — | 0.875 |
90 | noul | 7298 | 0.005 | 0.962 | 0.004 | 0.995 | 0.002 | — | — |
91 | score | 3343 | 0.030 | 0.872 | 0.003 | — | — | 0.116 | — |
92
93 ### test 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 | 457 | 0.238 | 0.860 | 0.018 | — | — | — | 0.860 |
98 | openjev_v2 | noul | 1139 | 0.012 | 0.994 | 0.004 | 1.000 | 0.004 | — | — |
99 | yuri_v1 | noul | 2951 | 0.000 | — | 0.009 | — | 0.000 | — | — |
100 | yuri_v3 | choice | 3163 | 0.037 | 0.877 | 0.003 | — | — | — | 0.877 |
101 | yuri_v3 | noul | 3208 | 0.006 | 0.951 | 0.003 | 0.992 | 0.003 | — | — |
102 | yuri_v3 | score | 3343 | 0.030 | 0.872 | 0.003 | — | — | 0.116 | — |
103
104 ## ood
105
106 all rows: `n=13058 · kl=0.2080 · js=0.0454 · top1=0.9164 · ece=0.0288 · mce=0.3191 · noul_auroc=0.9852 · noul_brier_soft=0.0427 · noul_brier_hard=0.0427 · score_mae_expected=0.7871 · score_rps=0.5324 · choice_top1=0.8397 · choice_kl=0.2989`
107
108 excluding yuri_v1 placeholders (0 rows): `n=13058 · kl=0.2080 · js=0.0454 · top1=0.9164 · ece=0.0288 · mce=0.3191 · noul_auroc=0.9852 · noul_brier_soft=0.0427 · noul_brier_hard=0.0427 · score_mae_expected=0.7871 · score_rps=0.5324 · choice_top1=0.8397 · choice_kl=0.2989`
109
110 throughput: 27200 tok/s · 102.6 rows/s
111
112 ### ood by kind
113
114 | kind | n | kl | top1 | ece | noul_auroc | noul_brier_soft | score_mae_expected | choice_top1 |
115 |---|---|---|---|---|---|---|---|---|
116 | choice | 3219 | 0.299 | 0.840 | 0.044 | — | — | — | 0.840 |
117 | noul | 9767 | 0.170 | 0.945 | 0.024 | 0.985 | 0.043 | — | — |
118 | score | 72 | 1.278 | 0.444 | 0.236 | — | — | 0.787 | — |
119
120 ### ood by source × kind
121
122 | source | kind | n | kl | top1 | ece | noul_auroc | noul_brier_soft | score_mae_expected | choice_top1 |
123 |---|---|---|---|---|---|---|---|---|---|
124 | openjev_v2 | choice | 3219 | 0.299 | 0.840 | 0.044 | — | — | — | 0.840 |
125 | openjev_v2 | noul | 9767 | 0.170 | 0.945 | 0.024 | 0.985 | 0.043 | — | — |
126 | openjev_v2 | score | 72 | 1.278 | 0.444 | 0.236 | — | — | 0.787 | — |
127
128 ## validation
129
130 all rows: `n=14111 · kl=0.0260 · js=0.0067 · top1=0.9111 · ece=0.0030 · mce=0.0097 · noul_auroc=0.9943 · noul_brier_soft=0.0019 · noul_brier_hard=0.0656 · score_mae_expected=0.1196 · score_rps=0.0115 · choice_top1=0.8843 · choice_kl=0.0644`
131
132 excluding yuri_v1 placeholders (2920 rows): `n=11191 · kl=0.0327 · js=0.0084 · top1=0.9111 · ece=0.0020 · mce=0.0097 · noul_auroc=0.9943 · noul_brier_soft=0.0031 · noul_brier_hard=0.0656 · score_mae_expected=0.1196 · score_rps=0.0115 · choice_top1=0.8843 · choice_kl=0.0644`
133
134 throughput: 26303 tok/s · 204.2 rows/s
135
136 ### validation by kind
137
138 | kind | n | kl | top1 | ece | noul_auroc | noul_brier_soft | score_mae_expected | choice_top1 |
139 |---|---|---|---|---|---|---|---|---|
140 | choice | 3592 | 0.064 | 0.884 | 0.005 | — | — | — | 0.884 |
141 | noul | 7318 | 0.005 | 0.963 | 0.004 | 0.994 | 0.002 | — | — |
142 | score | 3201 | 0.030 | 0.870 | 0.003 | — | — | 0.120 | — |
143
144 ### validation by source × kind
145
146 | source | kind | n | kl | top1 | ece | noul_auroc | noul_brier_soft | score_mae_expected | choice_top1 |
147 |---|---|---|---|---|---|---|---|---|---|
148 | openjev_v2 | choice | 463 | 0.239 | 0.859 | 0.027 | — | — | — | 0.859 |
149 | openjev_v2 | noul | 1167 | 0.015 | 0.992 | 0.005 | 1.000 | 0.005 | — | — |
150 | yuri_v1 | noul | 2920 | 0.000 | — | 0.009 | — | 0.000 | — | — |
151 | yuri_v3 | choice | 3129 | 0.039 | 0.888 | 0.003 | — | — | — | 0.888 |
152 | yuri_v3 | noul | 3231 | 0.006 | 0.952 | 0.002 | 0.992 | 0.003 | — | — |
153 | yuri_v3 | score | 3201 | 0.030 | 0.870 | 0.003 | — | — | 0.120 | — |
154
155 ## Choice permutation consistency
156
157 `{"n_rows": 1000, "mean_max_abs_diff": 0.030732551589608192, "p90_max_abs_diff": 0.06812191009521484, "top1_flip_rate": 0.04675}`
158