Has anyone tried RLHF on the GPT models when fine-tuning? Would this be useful?