ibm-research/POBS

Preference, Opinion, and Belief Survey (POBS)是一个综合性的评估大型语言模型主观倾向的基准数据集，包括偏好、观点和信仰，跨越有争议的、社会的、文化的和个人领域。该数据集由IBM研发，能够系统地、无需参考基准地评估LLM的主观性，帮助组织、研究人员和实践者在部署前审计模型的中立性、一致性和隐含偏见。

The Preference, Opinion, and Belief Survey (POBS) is a comprehensive benchmark for evaluating the subjective inclinations of Large Language Models (LLMs)—including preferences, opinions, and beliefs—across controversial, societal, cultural, and personal domains. Developed by IBM Research, POBS enables systematic and reference-free assessment of LLM subjectivity, helping organizations, researchers, and practitioners audit model neutrality, consistency, and implicit bias before deployment.