“Reinforcement Learning from Human Feedback (RLHF) provided a performance improvement equivalent to a 100x larger model size on human preference ratings in the StructGPT paper.”