That doesn't look like any definition of RLHF I'm familiar with, which amounts to first training a reward model on human feedback and then using that reward mod

That doesn't look like any definition of RLHF I'm familiar with, which amounts to first training a reward model on human feedback and then using that reward model to train the target model [1]. No reaching out to humans involved. [1] https://en.wikipedia.org/wiki/Reinforcement_learning_from_human_feedback https://en.wikipedia.org/wiki/Reinforcement_learning_from_hu...