WinoQueer
收藏资源简介:
WinoQueer是由南加州大学信息科学研究所开发的一个社区驱动的基准数据集,专门设计来检测大型语言模型中对LGBTQ+社区有害的偏见。该数据集通过社区调查收集,包含45540对句子,覆盖11种模板、9种性别身份群体、3种代词、60个常见名字和182个独特谓词。数据集的创建过程涉及广泛的社区参与,确保了数据集与实际社区需求的紧密联系。WinoQueer的应用领域主要集中在评估和减少大型语言模型中的反LGBTQ+偏见,以促进更公平和包容的技术发展。
WinoQueer is a community-driven benchmark dataset developed by the Information Sciences Institute of the University of Southern California, specifically designed to detect harmful biases against the LGBTQ+ community in large language models. Collected via community surveys, this dataset contains 45,540 sentence pairs, covering 11 templates, 9 gender identity groups, 3 pronouns, 60 common given names, and 182 unique predicates. The dataset's creation process involved extensive community participation, ensuring that it is closely aligned with the actual needs of the LGBTQ+ community. The primary application domains of WinoQueer focus on evaluating and mitigating anti-LGBTQ+ biases in large language models, so as to foster more equitable and inclusive technological development.




