562 B
562 B
| 1 | rank | experiment | eval_loss | train_loss | train_runtime |
|---|---|---|---|---|---|
| 2 | 1 | 10_r16_a32_lr1e4 | 0.6918 | 0.8612 | 7:15:07.38 |
| 3 | 2 | 06_lr1e4_r16_a16 | 0.7184 | 0.889 | 7:17:30.28 |
| 4 | 3 | 04_rank32_a32_lr5e5 | 0.7292 | 0.9003 | 7:22:32.83 |
| 5 | 4 | 05_r16_alpha32_lr5e5 | 0.7323 | 0.9026 | 7:17:43.75 |
| 6 | 5 | 02_optimizer_adamw8bit_r16_a16_lr5e5 | 0.7521 | 0.9248 | 7:20:23.21 |
| 7 | 6 | 09_batch2_accum4_r16_a16_lr5e5 | 0.7563 | 0.929 | 8:24:57.09 |
| 8 | 7 | 01_baseline_r16_a16_lr5e5_adamw | 0.7567 | 0.929 | 7:18:05.66 |
| 9 | 8 | 03_rank8_a16_lr5e5 | 0.7586 | 0.9308 | 7:16:23.07 |
| 10 | 9 | 07_lr2e5_r16_a16 | 0.7912 | 0.9674 | 7:18:24.50 |
| 11 | 10 | 08_attention_only_r16_a16_lr5e5 | 0.8094 | 0.9808 | 6:51:24.56 |