OffsetBias: Leveraging Debiased Data for Tuning Evaluators

TL;DR

This paper identifies biases in LLM-based evaluators, introduces a benchmark and de-biasing datasets, and demonstrates that fine-tuning on these datasets improves evaluator robustness and performance.

Contribution

It presents a comprehensive analysis of biases in LLM evaluators, introduces EvalBiasBench and OffsetBias datasets, and shows that fine-tuning on these datasets reduces bias effects.

Findings

Fine-tuning on OffsetBias reduces evaluator bias.

Robustness of judge models improves across evaluation scenarios.

Datasets and models are publicly released.

Abstract

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 meta-evaluation collection of hand-crafted test cases for each bias type. Additionally, we present de-biasing 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 fine-tuned judge model to public.