Jakub
67061cf373
Document six completed LR runs as preparation for full Aya training on Perun. Record selection of LR 0.00010 by lowest final validation loss, selected-model lm-eval and partial SK-bench results, reproducible provenance, and dataset-access/context-truncation limitations. Preserve existing 30k and Titan material.
1.2 KiB
1.2 KiB
| 1 | learning_rate | global_step | final_eval_step | final_eval_loss | min_eval_step | min_eval_loss | max_grad_norm | p95_grad_norm | trainer_state |
|---|---|---|---|---|---|---|---|---|---|
| 2 | 0.00008 | 18688 | 18688 | 0.875216543674469 | 18688 | 0.875216543674469 | 9.149866104125977 | 1.5837211608886719 | /home/schwarc/diplomovka/aya_300k_lr_validation/lr_0p00008/trainer_state.json |
| 3 | 0.00010 | 18688 | 18688 | 0.87261962890625 | 18688 | 0.87261962890625 | 8.578277587890625 | 1.6475107669830322 | /home/schwarc/diplomovka/aya_300k_lr_validation/lr_0p00010/trainer_state.json |
| 4 | 0.00015 | 18688 | 18688 | 0.8842697143554688 | 18688 | 0.8842697143554688 | 28.051427841186523 | 1.9590225219726562 | /home/schwarc/diplomovka/aya_300k_lr_validation/lr_0p00015/trainer_state.json |
| 5 | 0.00018 | 18688 | 18688 | 0.8914602994918823 | 18000 | 0.891387939453125 | 6.5584211349487305 | 2.163313627243042 | /home/schwarc/diplomovka/aya_300k_lr_validation/lr_0p00018/trainer_state.json |
| 6 | 0.00020 | 18688 | 18688 | 0.8976690173149109 | 18000 | 0.897506594657898 | 10.969125747680664 | 2.306173801422119 | /home/schwarc/diplomovka/aya_300k_lr_validation/lr_0p00020/trainer_state.json |
| 7 | 0.00023 | 18688 | 18688 | 0.9032362699508667 | 18000 | 0.9031931161880493 | 23.10581398010254 | 2.501060724258423 | /home/schwarc/diplomovka/aya_300k_lr_validation/lr_0p00023/trainer_state.json |