OffsetBias: Leveraging Debiased Data for Tuning Evaluators
Junsoo Park, Seungyeon Jwa, Ren Meiying, Daeyoung Kim, Sanghyuk Choi · 2024
Employing Large Language Models (LLMs) to assess the quality of generated responses, such as prompting instruct-tuned models or fine-tuning judge models, has become a widely adopted evaluation method.It is also known that such evaluators are vulnerable to biases, such as favoring longer responses.While it is important to overcome this problem, the specifics of these biases remain under-explored.In this work, we qualitatively identify six types of biases inherent in various judge models.We propose EVALBIASBENCH as a metaevaluation collection of hand-crafted test cases for each bias type.Additionally, we present debiasing dataset construction methods and the associated preference dataset OFFSETBIAS.Experimental results demonstrate that fine-tuning on our dataset significantly enhances the robustness of judge models against biases and improves performance across most evaluation scenarios.We release our datasets and the finetuned judge model to public. 1