last/train_state.json

{"step": 1000, "best_val_loss": 4.8720598220825195}