jasonhwan/iCliniq500-pdbd-race

该数据集是Persona-Aware Bias Evaluation (PABE)框架在医疗问答领域的一个小规模概念验证。数据集基于iCliniq-10K数据集，包含真实患者与医生的对话。通过对500条对话应用PABE框架，生成了2000条数据，每条数据包含原始输入、目标输出、增强后的输入和用于增强的人物属性。数据集旨在评估和减轻大型语言模型（LLM）中的偏见，特别是通过零样本文本风格转换和人物提示来隐式编码受保护属性信息。

This dataset is a small proof of concept (PoC) of the Persona-Aware Bias Evaluation (PABE) framework applied in the medical QA domain. The dataset is based on the iCliniq-10K dataset, which contains real-life patient-physician conversations. By applying the PABE framework to 500 conversations, a 2000-row dataset was generated, with each row containing the original input, target output, augmented input, and the persona attribute used for augmentation. The dataset aims to evaluate and mitigate bias in large language models (LLMs), particularly by implicitly encoding protected attribute information through zero-shot text style transfer and persona prompting.