NASK-PIB/RefusEU

这是一个多语言红队测试偏好和评估数据集，专门用于安全研究。它包含多语言红队测试提示和偏好对，以及一个独立的多语言评估提示集。数据集结构分为多个配置：multilingual配置提供完整的偏好训练和测试分割；lang_*配置（如lang_pl、lang_en、lang_cs等）提供特定语言的偏好训练和测试分割，覆盖波兰语、英语、捷克语、斯洛伐克语、斯洛文尼亚语、立陶宛语、拉脱维亚语、德语、意大利语、法语、西班牙语和葡萄牙语共12种语言；evaluation配置提供多语言评估提示。数据字段包括：偏好样本中的chosen（首选响应，以聊天轮次列表形式表示）、rejected（非首选响应，同样为聊天轮次列表）、lang（语言代码）、id（样本标识符）、row_id（原始行ID）、category（类别标签）；评估配置中的row_id、lang和prompt（评估提示文本）。该数据集旨在用于安全对齐和偏好学习研究、多语言拒绝和鲁棒性基准测试，以及在潜在有害提示分布下的模型评估。注意：数据集包含有害、滥用和违反政策的用户提示，部分被拒绝的助手响应可能包含不安全内容模式，需在受控研究环境中谨慎使用。

This dataset contains multilingual red-teaming prompts and preference pairs for safety research, plus a separate multilingual evaluation prompt set. The repository is organized into multiple dataset configs: multilingual config with full preference training and test splits; lang_* configs (e.g., lang_pl, lang_en, lang_cs) with language-specific preference training and test splits covering Polish, English, Czech, Slovak, Slovenian, Lithuanian, Latvian, German, Italian, French, Spanish, and Portuguese; and evaluation config with multilingual evaluation prompts. Data fields include: for preference samples, chosen (preferred response as a list of chat turns), rejected (non-preferred response as a list of chat turns), lang (language code), id (sample id), row_id (original row id), category (category label); for evaluation config, row_id, lang, prompt (evaluation prompt text). It is intended for safety alignment and preference learning research, multilingual refusal and robustness benchmarking, and model evaluation under potentially harmful prompt distributions. Safety notice: The dataset contains harmful, abusive, and policy-violating user prompts, and some rejected assistant responses may include unsafe content patterns; handle with care and use only in controlled research environments.