Have you tested whether the model's performance on 'correcting wrong answers to right ones' changes after SFT? Is the model just learning to be more confident in its answers?"