“The StructGPT paper demonstrated that using RLHF provides a performance improvement on human preference ratings equivalent to a 100x increase in model size compared to a non-RLHF model.”