saltlux/EthicsAI-B11-AugMT

EthicsAI-B11-AugMT是一个设计用于评估大型语言模型（LLM）在对话语境中识别和分析隐藏的社会偏见和刻板印象能力的数据集。该数据集基于著名的偏见检测数据集BBQ，扩展了约31,000个多轮对话场景。它不仅判断偏见的存在与否，还提供了偏见判断的逻辑理由和应对发言，以全面测量模型的伦理推理能力。数据集覆盖11种敏感主题，包括种族、宗教、社会经济地位、性别等，并通过多轮对话捕捉语境中的偏见。数据集中包含英语（27,702项）和韩语（3,639项），总规模为31,341项。

EthicsAI-B11-AugMT is a dataset designed to evaluate how accurately large language models (LLMs) can identify and analyze various social biases and stereotypes hidden in conversational contexts. It extends the famous bias detection dataset BBQ to approximately 31,000 multi-turn dialogue scenarios. The dataset goes beyond simply judging the presence of bias, including **logical reasons (Reason)** for why a particular statement is biased and **counter-utterances** to mitigate it, comprehensively measuring the models ethical reasoning capabilities. It covers 11 sensitive topics, including race, religion, socioeconomic status, gender, etc., and captures context-dependent biases through multi-turn dialogues. The dataset includes English (27,702 items) and Korean (3,639 items), with a total size of 31,341 items.