“Using more reward model data in RLHF leads to higher gold reward scores and reduces "Goodharting".”