qwen3-4b-think/train_results.json

{
    "epoch": 3.0,
    "total_flos": 1274755977576448.0,
    "train_loss": 0.5495063810720958,
    "train_runtime": 34913.8374,
    "train_samples_per_second": 3.33,
    "train_steps_per_second": 0.026
}