“Grouped with original ChatGPT as models trained using base model + supervised finetuning + reward model + RL/optimization.”