best/train_state.json

{"step": 1035, "best_val_loss": 4.8648576736450195}