TukaBench
收藏资源简介:
TukaBench是由Mila - 魁北克人工智能研究所、麦吉尔大学和微软人工智能公益研究实验室联合创建的针对七种非洲语言的越狱评估基准,旨在填补大语言模型安全评估在低资源语言领域的空白。该数据集包含986条提示词,涵盖阿姆哈拉语、豪萨语、伊博语等七种语言,通过人工翻译、文化适应和代码转换四种策略构建,数据来源于JailbreakBench的扩展和本地化创作。其创建过程采用机器翻译与人工后编辑相结合的三阶段流程,确保语言准确性和文化相关性,主要应用于评估大语言模型在非洲语言环境下的安全漏洞,解决因文化背景和语言资源不足导致的模型安全机制失效问题。
TukaBench is a jailbreak evaluation benchmark jointly developed by Mila - Quebec AI Institute, McGill University, and Microsoft AI for Good Research Lab, designed to fill the critical gap in large language model (LLM) safety assessment for low-resource African language domains. This dataset comprises 986 prompts across seven languages including Amharic, Hausa, Igbo and others, constructed via four strategies including manual translation, cultural adaptation and code-switching, with its corpus derived from the expansion and localization of the JailbreakBench dataset. The development of TukaBench follows a three-stage workflow integrating machine translation and human post-editing, ensuring both linguistic accuracy and cultural appropriateness. This benchmark is primarily utilized to assess the safety vulnerabilities of LLMs in African language contexts, addressing the failure of model safety mechanisms caused by insufficient cultural background and limited language resources.
数据集概述
- 数据集名称:Tukabench
- 所属机构:McGill-NLP(麦吉尔大学自然语言处理实验室)
- 共享平台:Hugging Face Datasets
- 访问地址:https://huggingface.co/datasets/McGill-NLP/tukabench
- 许可证类型:Apache-2.0
该数据集以 Apache-2.0 许可证开放使用,允许用户自由使用、修改和分发,但需保留版权声明和免责声明。当前页面提供的 README 内容仅包含许可证信息,未提供数据集的详细描述、任务类型、数据规模或具体用途说明。如需获取更完整的数据集信息,建议直接访问数据集页面或参考相关研究文献。




