一字一诗 / 第 16 课 / 图 08
正式训练存档 · 不是本次短跑

正式模型:18000 步的完整训练记录

finished 是结束状态;step=18000 是完成更新数;parameters=4877312 是正式模型大小。

run.json 原始字段节选

{
  "status": "finished",
  "step": 18000,
  "parameters": 4877312,
  "device": "mps",
  "seconds": 1200.65,
  "best_val": 4.084603522531118
}

metrics.jsonl 原始记录节选:0 / 6000 / 18000 步

{"event": "baseline", "step": 0, "val_loss": 8.86757787443655}
{"event": "validation", "step": 6000, "val_loss": 4.315424796221505, "best_val": 4.315424796221505, "peak_process_rss_mib": 600.8, "mps_driver_allocated_mib": 1217.5}
{"event": "validation", "step": 18000, "val_loss": 4.084603522531118, "best_val": 4.084603522531118, "peak_process_rss_mib": 704.5, "mps_driver_allocated_mib": 1281.5}