Reinforcement Learning from Human Feedback (RLHF) has become a dominating strategy in aligning Language Models (LMs) with human values/goals. The key to the strategy is learning a reward model ($\varphi$), which can reflect the latent reward model of humans. While this strategy has proven effective, the training methodology requires a lot of human preference annotation (usually in the order of tens of thousands) to train $\varphi$. Such a large-scale annotation is justifiable when it's a one-time effort, and the reward model is universally applicable. However, human goals are subjective and depend on the task, requiring task-specific preference annotations, which can be impractical to fulfill. To address this challenge, we propose a novel approach to infuse domain knowledge into $\varphi$, which reduces the amount of preference annotation required ($21\times$), omits Alignment Tax, and provides some interpretability. We validate our approach in E-Commerce Opinion Summarization, with a significant reduction in dataset size (to just $940$ samples) while advancing the SOTA ($\sim4$ point ROUGE-L improvement, $68\%$ of times preferred by humans over SOTA). Our contributions include a novel Reward Modeling technique and two new datasets: PromptOpinSumm (supervised data for Opinion Summarization) and OpinPref (a gold-standard human preference dataset). The proposed methodology opens up avenues for efficient RLHF, making it more adaptable to applications with varying human values. We release the artifacts (Code: github.com/efficient-rlhf. PromptOpinSumm: hf.co/prompt-opin-summ. OpinPref: hf.co/opin-pref) for usage under MIT License.
翻译:暂无翻译