启动一次 50 步的真实短训练

第一行 start_step=0；baseline 是还没学习时的成绩；最后看 validation 的 step=50。

实际执行的命令

./poet train --config configs/tiny.json --run-dir artifacts/runs/illustrated-demo --steps 50 --batch-size 24 --device mps

标准输出

{"device": "mps", "parameters": 1453952, "start_step": 0, "target_steps": 50, "train_poems": 87236}
{"event": "baseline", "step": 0, "val_loss": 8.826677150827237}
{"event": "train", "step": 1, "loss": 8.833823204040527, "lr": 0.00014000000000000001, "seconds": 0.77, "steps_per_second": 1.296}
{"event": "train", "step": 25, "loss": 7.221395969390869, "lr": 0.0004541603328559118, "seconds": 1.25, "steps_per_second": 20.039}
{"event": "train", "step": 50, "loss": 6.819660186767578, "lr": 0.00010499999999999999, "seconds": 1.71, "steps_per_second": 29.257}
{"event": "validation", "step": 50, "val_loss": 6.8024335038725035, "best_val": 6.8024335038725035, "peak_process_rss_mib": 545.6, "mps_driver_allocated_mib": 1166.2}

执行结果

退出码 0 · 2026-09-19T00:55:47.402201+00:00