“Surge holds a strong view that Reinforcement Learning from Human Feedback (RLHF) data is significantly more effective for training LLMs than Supervised Fine-Tuning (SFT) data.”