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.
93 lines
2.5 KiB
JSON
93 lines
2.5 KiB
JSON
[
|
|
{
|
|
"learning_rate": 8e-05,
|
|
"run": "lr_0p00008",
|
|
"global_step": 18688,
|
|
"epoch": 1.0,
|
|
"final_eval_step": 18688,
|
|
"final_eval_loss": 0.875216543674469,
|
|
"min_eval_step": 18688,
|
|
"min_eval_loss": 0.875216543674469,
|
|
"max_grad_norm": 9.149866104125977,
|
|
"p95_grad_norm": 1.5837211608886719,
|
|
"train_runtime": 79243.6449,
|
|
"eval_observations": 19,
|
|
"gradient_observations": 1868
|
|
},
|
|
{
|
|
"learning_rate": 0.0001,
|
|
"run": "lr_0p00010",
|
|
"global_step": 18688,
|
|
"epoch": 1.0,
|
|
"final_eval_step": 18688,
|
|
"final_eval_loss": 0.87261962890625,
|
|
"min_eval_step": 18688,
|
|
"min_eval_loss": 0.87261962890625,
|
|
"max_grad_norm": 8.578277587890625,
|
|
"p95_grad_norm": 1.6475107669830322,
|
|
"train_runtime": 79344.5851,
|
|
"eval_observations": 19,
|
|
"gradient_observations": 1868
|
|
},
|
|
{
|
|
"learning_rate": 0.00015,
|
|
"run": "lr_0p00015",
|
|
"global_step": 18688,
|
|
"epoch": 1.0,
|
|
"final_eval_step": 18688,
|
|
"final_eval_loss": 0.8842697143554688,
|
|
"min_eval_step": 18688,
|
|
"min_eval_loss": 0.8842697143554688,
|
|
"max_grad_norm": 28.051427841186523,
|
|
"p95_grad_norm": 1.9590225219726562,
|
|
"train_runtime": 80127.6428,
|
|
"eval_observations": 19,
|
|
"gradient_observations": 1868
|
|
},
|
|
{
|
|
"learning_rate": 0.00018,
|
|
"run": "lr_0p00018",
|
|
"global_step": 18688,
|
|
"epoch": 1.0,
|
|
"final_eval_step": 18688,
|
|
"final_eval_loss": 0.8914602994918823,
|
|
"min_eval_step": 18000,
|
|
"min_eval_loss": 0.891387939453125,
|
|
"max_grad_norm": 6.5584211349487305,
|
|
"p95_grad_norm": 2.163313627243042,
|
|
"train_runtime": 79892.0077,
|
|
"eval_observations": 19,
|
|
"gradient_observations": 1868
|
|
},
|
|
{
|
|
"learning_rate": 0.0002,
|
|
"run": "lr_0p00020",
|
|
"global_step": 18688,
|
|
"epoch": 1.0,
|
|
"final_eval_step": 18688,
|
|
"final_eval_loss": 0.8976690173149109,
|
|
"min_eval_step": 18000,
|
|
"min_eval_loss": 0.897506594657898,
|
|
"max_grad_norm": 10.969125747680664,
|
|
"p95_grad_norm": 2.306173801422119,
|
|
"train_runtime": 79710.3252,
|
|
"eval_observations": 19,
|
|
"gradient_observations": 1868
|
|
},
|
|
{
|
|
"learning_rate": 0.00023,
|
|
"run": "lr_0p00023",
|
|
"global_step": 18688,
|
|
"epoch": 1.0,
|
|
"final_eval_step": 18688,
|
|
"final_eval_loss": 0.9032362699508667,
|
|
"min_eval_step": 18000,
|
|
"min_eval_loss": 0.9031931161880493,
|
|
"max_grad_norm": 23.10581398010254,
|
|
"p95_grad_norm": 2.501060724258423,
|
|
"train_runtime": 79789.3562,
|
|
"eval_observations": 19,
|
|
"gradient_observations": 1868
|
|
}
|
|
]
|