SB-Bench
收藏资源简介:
SB-Bench是一个用于评估大型多模态模型中刻板印象偏见的全面基准,由中央佛罗里达大学的研究团队开发。该数据集包含7500个真实图像的三元组,涵盖了年龄、残疾状况、性别认同、国籍、种族/ ethnicity、宗教、性取向、外貌和社会经济地位等九个多样化的领域,及其60个子类别。这些图像与情境信息和多项选择题相结合,为评估多模态模型中的视觉刻板印象提供了精确和细致的手段。
SB-Bench is a comprehensive benchmark for evaluating stereotypical biases in large multimodal models, developed by a research team from the University of Central Florida. This dataset contains 7,500 real-image triplets, covering nine diverse domains including age, disability status, gender identity, nationality, race/ethnicity, religion, sexual orientation, physical appearance, and socioeconomic status, which encompass 60 subcategories in total. These images are combined with contextual information and multiple-choice questions, providing a precise and meticulous approach to evaluating visual stereotypes in multimodal models.

- 1SB-Bench: Stereotype Bias Benchmark for Large Multimodal Models中央佛罗里达大学 · 2025年



