3 ms·
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 a
by piecerough 2y 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