checkpoints/epoch_16/training_state.json

{"epoch": 16, "average_batch_loss": 11.65863088607788}