8a73394df9
Re-run pre-training results with the corrected 84-token vocabulary and max_seq_len=320. Previous checkpoint was trained on stale data with BAR tokens and a corrupted tokenizer. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
68 lines
3.0 KiB
Plaintext
68 lines
3.0 KiB
Plaintext
|
|
====================================================
|
|
PRE-TRAINING REPORT
|
|
====================================================
|
|
Total epochs run : 50
|
|
Best epoch (val loss) : 49
|
|
Convergence epoch : 42 (val ≤ best+1 %)
|
|
Best val loss : 0.2785
|
|
Best val perplexity : 1.32
|
|
Final train loss : 0.2539
|
|
Unique parameters : 1,396,416
|
|
Checkpoint : checkpoints/pretrained.pt
|
|
Log CSV : checkpoints/pretrained.log.csv
|
|
====================================================
|
|
|
|
epoch train val ppl lr
|
|
----- -------- -------- ------- ----------
|
|
1 2.0431 0.8604 2.36 2.20e-04
|
|
2 0.6824 0.5873 1.80 3.00e-04
|
|
3 0.5679 0.5449 1.72 2.99e-04
|
|
4 0.5294 0.5129 1.67 2.98e-04
|
|
5 0.5054 0.4908 1.63 2.96e-04
|
|
6 0.4849 0.4717 1.60 2.93e-04
|
|
7 0.4671 0.4569 1.58 2.90e-04
|
|
8 0.4502 0.4428 1.56 2.86e-04
|
|
9 0.4359 0.4285 1.53 2.82e-04
|
|
10 0.4256 0.4201 1.52 2.77e-04
|
|
11 0.4148 0.4112 1.51 2.72e-04
|
|
12 0.4055 0.4097 1.51 2.66e-04
|
|
13 0.3969 0.3919 1.48 2.60e-04
|
|
14 0.3876 0.3873 1.47 2.53e-04
|
|
15 0.3791 0.3851 1.47 2.45e-04
|
|
16 0.3717 0.3745 1.45 2.38e-04
|
|
17 0.3645 0.3673 1.44 2.30e-04
|
|
18 0.3574 0.3645 1.44 2.21e-04
|
|
19 0.3503 0.3585 1.43 2.13e-04
|
|
20 0.3430 0.3498 1.42 2.04e-04
|
|
21 0.3377 0.3438 1.41 1.95e-04
|
|
22 0.3308 0.3370 1.40 1.85e-04
|
|
23 0.3248 0.3323 1.39 1.76e-04
|
|
24 0.3194 0.3249 1.38 1.66e-04
|
|
25 0.3141 0.3215 1.38 1.57e-04
|
|
26 0.3098 0.3177 1.37 1.47e-04
|
|
27 0.3043 0.3134 1.37 1.37e-04
|
|
28 0.3000 0.3108 1.36 1.28e-04
|
|
29 0.2950 0.3072 1.36 1.18e-04
|
|
30 0.2901 0.3034 1.35 1.09e-04
|
|
31 0.2880 0.3020 1.35 9.95e-05
|
|
32 0.2835 0.2993 1.35 9.05e-05
|
|
33 0.2805 0.2948 1.34 8.17e-05
|
|
34 0.2759 0.2919 1.34 7.32e-05
|
|
35 0.2737 0.2888 1.33 6.51e-05
|
|
36 0.2706 0.2878 1.33 5.73e-05
|
|
37 0.2679 0.2865 1.33 4.98e-05
|
|
38 0.2660 0.2848 1.33 4.28e-05
|
|
39 0.2645 0.2837 1.33 3.63e-05
|
|
40 0.2623 0.2827 1.33 3.02e-05
|
|
41 0.2608 0.2822 1.33 2.46e-05
|
|
42 0.2589 0.2807 1.32 1.96e-05
|
|
43 0.2579 0.2802 1.32 1.51e-05
|
|
44 0.2568 0.2794 1.32 1.11e-05
|
|
45 0.2549 0.2793 1.32 7.75e-06
|
|
46 0.2556 0.2789 1.32 4.98e-06
|
|
47 0.2550 0.2787 1.32 2.81e-06
|
|
48 0.2543 0.2786 1.32 1.25e-06
|
|
49 0.2524 0.2785 1.32 3.13e-07 ←
|
|
50 0.2539 0.2785 1.32 0.00e+00
|