VL-Bias

该数据集包含24,000组图像-文本对，旨在衡量视觉-语言预训练模型中的性别偏见。该数据集创新性地构建了包含性别、活动以及职业概念的标注，以便深入研究视觉-语言预训练模型在社会偏见分析方面的表现。具体来说，数据集中的图像-文本对分为13,000组活动相关和11,000组职业相关，这些详细的数据规模和任务设计有助于分析视觉-语言预训练模型中的性别偏见问题。This dataset contains 24,000 image-text pairs, aimed at measuring gender bias in vision-language pre-trained models. It innovatively constructs annotations covering the concepts of gender, activities and occupations, to enable in-depth research on the performance of vision-language pre-trained models in social bias analysis. Specifically, the image-text pairs in the dataset are divided into 13,000 activity-related pairs and 11,000 occupation-related pairs. Such detailed data scale and task design facilitate the analysis of gender bias in vision-language pre-trained models.