Cheems: A Practical Guidance for Building and Evaluating Chinese Reward Models from Scratch

TL;DR

This paper introduces CheemsBench and CheemsPreference to improve Chinese reward models, highlighting the importance of human supervision for better alignment with human preferences in Chinese language tasks.

Contribution

It provides the first Chinese-specific RM evaluation benchmark and a large-scale preference dataset, demonstrating the effectiveness of human supervision in training high-quality Chinese reward models.

Findings

Open-source RMs have limited ability to capture Chinese human preferences.

Human-supervised training yields state-of-the-art Chinese RM performance.

AI-generated data alone is insufficient for aligning models with human preferences.

Abstract

Reward models (RMs) are crucial for aligning large language models (LLMs) with human preferences. However, most RM research is centered on English and relies heavily on synthetic resources, which leads to limited and less reliable datasets and benchmarks for Chinese. To address this gap, we introduce CheemsBench, a fully human-annotated RM evaluation benchmark within Chinese contexts, and CheemsPreference, a large-scale and diverse preference dataset annotated through human-machine collaboration to support Chinese RM training. We systematically evaluate open-source discriminative and generative RMs on CheemsBench and observe significant limitations in their ability to capture human preferences in Chinese scenarios. Additionally, based on CheemsPreference, we construct an RM that achieves state-of-the-art performance on CheemsBench, demonstrating the necessity of human supervision in RM training. Our findings reveal that scaled AI-generated data struggles to fully capture human preferences, emphasizing the importance of high-quality human supervision in RM development.