政治偏见测量数据集
收藏资源简介:
该数据集是一个包含88,110条观察的政治偏见测量数据集,由康斯坦茨大学、曼海姆大学和巴塞罗那超级计算中心的研究人员创建。数据集通过三十种不同的提示变体,对十一种开源和商业生成型大型语言模型进行了政治偏见评估,旨在研究这些模型在处理政治声明时的立场和偏见。数据集的创建过程中,研究人员使用了世界价值观调查(WVS)和政冶指南针测试(PCT)中的政治声明,并利用GPT-4模型对声明进行改写和反转,以测试模型在不同政治立场上的反应。该数据集用于分析和测量大型语言模型的政治偏见,有助于解决模型在信息收集、搜索或内容分析相关任务中可能存在的政治偏见问题。
This is a political bias measurement dataset containing 88,110 observations, developed by researchers from the University of Konstanz, the University of Mannheim, and the Barcelona Supercomputing Center. It evaluated eleven open-source and commercial generative large language models using thirty distinct prompt variants, aiming to investigate the stances and biases of these models when processing political statements. During the dataset development process, researchers adopted political statements from the World Values Survey (WVS) and the Political Compass Test (PCT), and used the GPT-4 model to rewrite and invert these statements to test the models' responses across different political stances. This dataset is employed to analyze and measure political biases in large language models, helping to address potential political bias issues in tasks related to information gathering, search, or content analysis.




