Isn't distillation done through SFT by having the student model approximate the teacher's output distribution?