checkpoints/epoch_2/training_state.json

{"epoch": 2, "average_batch_loss": 43.480728454589844}