checkpoints/epoch_20/training_state.json

{"epoch": 20, "average_batch_loss": 11.100186767578125}