Training loss (should decrease smoothly) - Validation loss (should decrease; divergence from training loss indicates overfitting) - Response quality samples (manual inspection of generated responses)