CBBQ
收藏资源简介:
CBBQ是由天津大学的研究团队通过人工与AI协作精心构建的中文偏见基准数据集,包含超过10万个问题,覆盖了中国文化和社会价值观相关的14个社会偏见维度。数据集的创建过程包括偏见识别、模糊上下文生成、AI辅助的清晰上下文生成以及人工审查与重组四个关键步骤。CBBQ旨在为大型语言模型提供一个全面且有效的测试平台,以评估和减少模型中的伦理风险,特别是在预部署阶段。此外,数据集还展示了模型在特定类别中表现出的偏见,如教育资格、疾病、残疾和外貌等,同时揭示了模型在一定程度上能够遵循指令进行道德自我修正的可能性。
CBBQ is a meticulously constructed Chinese bias benchmark dataset developed by the research team from Tianjin University through human-AI collaboration. It comprises over 100,000 questions spanning 14 social bias dimensions related to Chinese culture and social values. The dataset's creation process involves four core steps: bias identification, ambiguous context generation, AI-assisted clear context generation, and manual review and reorganization. CBBQ aims to serve as a comprehensive and effective testbed for large language models (LLMs) to evaluate and mitigate ethical risks in such models, particularly during the pre-deployment phase. Additionally, the dataset demonstrates the biases exhibited by models in specific categories such as educational qualifications, diseases, disabilities, and physical appearance, while also revealing the potential that models can follow instructions to conduct moral self-correction to a certain extent.

- 1CBBQ: A Chinese Bias Benchmark Dataset Curated with Human-AI Collaboration for Large Language Models天津大学 · 2023年



