train_stats.json

{
 "base": "microsoft/phi-4",
 "method": "fsdp2-custom-loop",
 "n_rows": 22321,
 "n_self": 0,
 "n_sequences": 9508,
 "n_tokens": 15596130,
 "world": 4,
 "micro_batch": 1,
 "accum": 16,
 "steps": 148,
 "lr": 1e-05,
 "optim": "AdamW fused, fp32 master, bf16 compute, fp32 grad reduce",
 "loss_curve": [
  1.5430302619934082,
  1.4843451976776123,
  1.297351598739624,
  1.3100507259368896,
  1.2504206895828247,
  1.250080943107605,
  1.154129981994629,
  1.1947247982025146,
  1.24116849899292,
  1.2432225942611694,
  1.244370460510254,
  1.1873300075531006,
  1.1999728679656982,
  1.3423196077346802,
  1.23753821849823,
  1.3136370182037354,
  1.183390498161316,
  1.1182904243469238,
  1.2329813241958618,
  1.2145954370498657,
  1.2024600505828857,
  1.1313198804855347,
  1.1699527502059937,
  1.1228104829788208,
  1.185697317123413,
  1.2165992259979248,
  1.2076386213302612,
  1.157266616821289,
  1.2494041919708252,
  1.1573677062988281
 ],
 "loss_first": [
  1.5430302619934082
 ],
 "loss_last": [
  1.1573677062988281
 ],
 "train_seconds": 3397.1226258277893,
 "tokens_per_second": 4590.9823453016015,
 "resumed_from": null
}