checkpoints/epoch_12/training_state.json

{"epoch": 12, "average_batch_loss": 9.142969264984131}