| # Selected scientific lines; this is a derived compact log, not the full raw worker log. |
| 2026-10-19 03:21:53 [cfg] unit_bf16_d910_c1_lr0.001_acc1_s99_st152580_ev5000_corpc420b_perm_cos_mb8_w8 |
| 2026-10-19 03:21:53 [env] {'torch': '2.11.0+cu128', 'gpu': 'NVIDIA B200', 'tf32_matmul': False, 'f32_mode': 'highest', 'python': '3.12.3'} |
| 2026-10-19 03:22:01 [model] 967.0M parameters |
| 2026-10-19 03:22:01 [data] corpus c420b (c4_20b_tokens.i32) |
| 2026-10-19 03:22:01 [data] 19999.70M training tokens; this cell processes 19999.0M, 1.00 of the corpus |
| 2026-10-19 03:22:01 [ood] 256 sequences from wikitext_train_tokens.pt |
| 2026-10-19 03:22:02 [radius] max over 49 normalizers 30.150; u = R/sqrt(c) = 30.150 |
| 2026-10-19 03:22:02 [sampler] permutation of 9765477 windows; this cell uses 9765120 of them (1.000 passes) |
| 2026-10-19 03:22:02 [sched] warmup 3051 steps to 0.001, cosine to 0.0001 at step 152580 |
| 2026-10-19 03:41:34 [health] steps 1-1000: loss 6.9101 grad-norm mean 1.420 max 3.707 clipped 74% |
| 2026-10-19 03:41:44 [resume] saved at step 1000 |
| 2026-10-19 04:01:43 [health] steps 1001-2000: loss 4.9224 grad-norm mean 0.989 max 1.714 clipped 43% |
| 2026-10-19 04:01:56 [resume] saved at step 2000 |
| 2026-10-19 04:21:57 [health] steps 2001-3000: loss 4.3102 grad-norm mean 0.661 max 1.232 clipped 2% |
| 2026-10-19 04:22:11 [resume] saved at step 3000 |
| 2026-10-19 04:42:16 [health] steps 3001-4000: loss 4.0557 grad-norm mean 0.489 max 2.412 clipped 0% |
| 2026-10-19 04:42:29 [resume] saved at step 4000 |
| 2026-10-19 05:02:57 [eval] first evaluation did not change model state (d8f031bb8ffa6965 -> d8f031bb8ffa6965) |
| 2026-10-19 05:03:11 [unit_bf16_d910_c1_lr0.001_acc1_s99_st152580_ev5000_corpc420b_perm_cos_mb8_w8] step 5000 ppl 47.1 ood 111.5 (101m train) |
| 2026-10-19 05:03:11 [health] steps 4001-5000: loss 3.8883 grad-norm mean 0.448 max 3.254 clipped 1% |
| 2026-10-19 05:03:24 [resume] saved at step 5000 |
| 2026-10-19 05:23:31 [health] steps 5001-6000: loss 3.8033 grad-norm mean 0.400 max 1.610 clipped 0% |
| 2026-10-19 05:23:44 [resume] saved at step 6000 |
| 2026-10-19 05:43:47 [health] steps 6001-7000: loss 3.7367 grad-norm mean 0.421 max 3.393 clipped 2% |
| 2026-10-19 05:44:00 [resume] saved at step 7000 |
| 2026-10-19 06:04:00 [health] steps 7001-8000: loss 3.6892 grad-norm mean 0.429 max 4.317 clipped 2% |
| 2026-10-19 06:04:13 [resume] saved at step 8000 |
| 2026-10-19 06:24:11 [health] steps 8001-9000: loss 3.6558 grad-norm mean 0.451 max 10.302 clipped 3% |
| 2026-10-19 06:24:24 [resume] saved at step 9000 |
| 2026-10-19 06:44:49 [unit_bf16_d910_c1_lr0.001_acc1_s99_st152580_ev5000_corpc420b_perm_cos_mb8_w8] step 10000 ppl 37.0 ood 88.3 (202m train) |
| 2026-10-19 06:44:49 [health] steps 9001-10000: loss 3.6154 grad-norm mean 0.436 max 16.802 clipped 1% |
| 2026-10-19 06:45:02 [resume] saved at step 10000 |
| 2026-10-19 06:45:02 [stop] step limit at step 10000; evaluation and resume saved |
| 2026-10-19 06:45:02 [done] unit_bf16_d910_c1_lr0.001_acc1_s99_st152580_ev5000_corpc420b_perm_cos_mb8_w8: STEP-LIMITED at step 10000; evaluation and resume retained; planned horizon remains 152580 |
|
|
