Reward model

A model trained on human comparisons to give a score to an output, such as an answer. It stands in for human judgement during reinforcement learning, so people do not have to rate every output.