checkpoints/epoch_32/training_state.json

{"epoch": 32, "average_batch_loss": 10.844172954559326}