reports/data_audit.md

2.9 KB · 71 lines · markdown Raw

1	# Data audit — `jev-distill-corpus-v3`
2
3	tokenizer: `/root/models/Qwen3.5-9B` · max_seq_len: 1024 · template: bare-v1 (whole-string tokenization)
4
5	`## Splits`
6
7	`\| split \| rows \| schema violations \| mean tok \| p50 \| p95 \| p99 \| max \| > max_seq_len \| time \|`
8	`\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|`
9	`\| train \| 655,806 \| 0 \| 129.0 \| 87 \| 351 \| 531 \| 856 \| 0 \| 83s \|`
10	`\| validation \| 14,111 \| 0 \| 128.8 \| 87 \| 339 \| 530 \| 699 \| 0 \| 2s \|`
11	`\| calibration \| 13,766 \| 0 \| 128.2 \| 87 \| 354 \| 525 \| 699 \| 0 \| 2s \|`
12	`\| test \| 14,261 \| 0 \| 128.5 \| 87 \| 345 \| 527 \| 716 \| 0 \| 1s \|`
13	`\| test_set_30k \| 29,955 \| 0 \| 106.6 \| 85 \| 214 \| 509 \| 702 \| 0 \| 3s \|`
14	`\| ood \| 13,058 \| 0 \| 265.0 \| 178 \| 553 \| 662 \| 734 \| 0 \| 2s \|`
15
16	`train tokens / epoch ≈ 84.6M; 2 epochs ≈ 169M`
17
18	`## D1 — exactly-uniform teacher labels by split × source × kind`
19
20	`\| split \| source \| kind \| rows \| uniform \| share \|`
21	`\|---\|---\|---\|---\|---\|---\|`
22	`\| train \| openjev_v2 \| choice \| 20,434 \| 511 \| 2.5% \|`
23	`\| train \| openjev_v2 \| noul \| 54,192 \| 0 \| 0.0% \|`
24	`\| train \| yuri_v1 \| noul \| 137,203 \| 137,203 \| 100.0% \|`
25	`\| train \| yuri_v3 \| choice \| 143,223 \| 0 \| 0.0% \|`
26	`\| train \| yuri_v3 \| noul \| 150,515 \| 1,514 \| 1.0% \|`
27	`\| train \| yuri_v3 \| score \| 150,239 \| 0 \| 0.0% \|`
28	`\| validation \| openjev_v2 \| choice \| 463 \| 15 \| 3.2% \|`
29	`\| validation \| openjev_v2 \| noul \| 1,167 \| 0 \| 0.0% \|`
30	`\| validation \| yuri_v1 \| noul \| 2,920 \| 2,920 \| 100.0% \|`
31	`\| validation \| yuri_v3 \| choice \| 3,129 \| 0 \| 0.0% \|`
32	`\| validation \| yuri_v3 \| noul \| 3,231 \| 35 \| 1.1% \|`
33	`\| validation \| yuri_v3 \| score \| 3,201 \| 0 \| 0.0% \|`
34	`\| calibration \| openjev_v2 \| choice \| 463 \| 9 \| 1.9% \|`
35	`\| calibration \| openjev_v2 \| noul \| 1,109 \| 0 \| 0.0% \|`
36	`\| calibration \| yuri_v1 \| noul \| 2,812 \| 2,812 \| 100.0% \|`
37	`\| calibration \| yuri_v3 \| choice \| 2,955 \| 0 \| 0.0% \|`
38	`\| calibration \| yuri_v3 \| noul \| 3,254 \| 30 \| 0.9% \|`
39	`\| calibration \| yuri_v3 \| score \| 3,173 \| 0 \| 0.0% \|`
40	`\| test \| openjev_v2 \| choice \| 457 \| 8 \| 1.8% \|`
41	`\| test \| openjev_v2 \| noul \| 1,139 \| 0 \| 0.0% \|`
42	`\| test \| yuri_v1 \| noul \| 2,951 \| 2,951 \| 100.0% \|`
43	`\| test \| yuri_v3 \| choice \| 3,163 \| 0 \| 0.0% \|`
44	`\| test \| yuri_v3 \| noul \| 3,208 \| 37 \| 1.2% \|`
45	`\| test \| yuri_v3 \| score \| 3,343 \| 0 \| 0.0% \|`
46	`\| test_set_30k \| openjev_v2 \| choice \| 887 \| 29 \| 3.3% \|`
47	`\| test_set_30k \| openjev_v2 \| noul \| 1,432 \| 0 \| 0.0% \|`
48	`\| test_set_30k \| yuri_v1 \| noul \| 2,260 \| 2,260 \| 100.0% \|`
49	`\| test_set_30k \| yuri_v3 \| choice \| 8,312 \| 0 \| 0.0% \|`
50	`\| test_set_30k \| yuri_v3 \| noul \| 8,537 \| 91 \| 1.1% \|`
51	`\| test_set_30k \| yuri_v3 \| score \| 8,527 \| 0 \| 0.0% \|`
52	`\| ood \| openjev_v2 \| choice \| 3,219 \| 0 \| 0.0% \|`
53	`\| ood \| openjev_v2 \| noul \| 9,767 \| 0 \| 0.0% \|`
54	`\| ood \| openjev_v2 \| score \| 72 \| 0 \| 0.0% \|`
55
56	`## Train mix`
57
58	`\| kind \| rows \| share \|`
59	`\|---\|---\|---\|`
60	`\| noul \| 341,910 \| 52.1% \|`
61	`\| choice \| 163,657 \| 25.0% \|`
62	`\| score \| 150,239 \| 22.9% \|`
63
64	`\| source \| rows \| share \|`
65	`\|---\|---\|---\|`
66	`\| yuri_v3 \| 443,977 \| 67.7% \|`
67	`\| yuri_v1 \| 137,203 \| 20.9% \|`
68	`\| openjev_v2 \| 74,626 \| 11.4% \|`
69
70	`choice n_options histogram (train): 2:910, 3:8,826, 4:142,002, 5:3,586, 6:511, 7:1,593, 8:8, 9:3,121, 16:3,100`
71