Hacker News new | ask | show | jobs
by piecerough 505 days ago
SFT forces the model to output _that_ reasoning trace you have in data. RL allows whatever reasoning trace and only penalizes it if it does not reach the same answer