BiasLens
收藏资源简介:
BiasLens数据集由北京大学、南洋理工大学等机构的研究人员创建,旨在系统地检测大型语言模型(LLMs)在角色扮演场景中的社会偏见。该数据集包含33,000个问题,这些问题基于11个不同的社会属性生成,涵盖了多种问题格式,如Yes/No、多选和开放式问题。数据集的创建过程利用了LLMs生成角色和问题,并通过规则和LLM辅助策略识别偏见响应。BiasLens数据集主要应用于评估和改进LLMs在实际应用中的公平性,特别是在涉及角色扮演的任务中,旨在解决模型输出中的偏见问题。
The BiasLens dataset was developed by researchers from Peking University, Nanyang Technological University and other institutions, with the aim of systematically detecting social biases in large language models (LLMs) during role-playing scenarios. This dataset consists of 33,000 questions generated based on 11 distinct social attributes, covering multiple question formats including Yes/No, multiple-choice, and open-ended questions. The dataset construction process leverages LLMs to generate roles and questions, and identifies biased responses through rule-based and LLM-aided strategies. The BiasLens dataset is primarily applied to evaluate and enhance the fairness of LLMs in real-world applications, especially in tasks involving role-playing, with the goal of mitigating biases in model outputs.




