Adversarial Prompts and Counterfactuals

该数据集包含了专门设计用来引发目标大型语言模型（LLM）偏见响应的对抗性提示，以及旨在通过替换性别化词汇来减轻偏见的反事实提示。此外，该数据集还包含了用于测试不同LLM偏见程度的提示，并结合LLM作为评判工具，用于自动化偏见评估。该数据集是从更大规模的生成提示中筛选出的约500个提示的子集，其任务是对性别偏见进行检测和评估。

This dataset contains adversarial prompts specifically designed to elicit biased responses from target large language models (LLMs), as well as counterfactual prompts intended to mitigate bias by replacing gendered terms. Additionally, the dataset includes prompts for testing the degree of gender bias across different LLMs, paired with LLM-as-judge tools for automated bias evaluation. This dataset is a subset of approximately 500 prompts filtered from a larger corpus of generated prompts, focused on the task of gender bias detection and assessment.