Towards Comprehensive Preference Data Collection for Reward Modeling

Reinforcement Learning from Human Feedback (RLHF) facilitates the alignment of large language models (LLMs) with human preferences. A critical component of RLHF is the reward model, which is trained on preference data and outputs a scalar reward for given text. However, the collection of high-quality preference data still lacks thorough investigation. Recent studies indicate that preference data is collected either by AI or humans, where chosen and rejected instances are identified between pairwise responses. We question whether this process effectively filters out noise and ensures sufficient diversity in the collected data. To address these concerns, for the first time, we propose a comprehensive framework for preference data collection, decomposing the process into four incremental steps: Prompt Collection, Response Generation, Response Filtering, and Human Labeling. This framework ensures the collection of high-quality preferences while reducing reliance on human labor. We conducted comprehensive experiments using the data collected at different stages, demonstrating the effectiveness of the proposed data collection method.