DemOgraphic FActualIty Representation (DoFaiR)
收藏资源简介:
DemOgraphic FActualIty Representation (DoFaiR)数据集由加州大学洛杉矶分校创建,旨在评估文本到图像模型在多样性和事实性之间的平衡。该数据集包含756条记录,涉及不同的历史事件、参与者和相应的人口统计信息。数据集的创建过程包括从维基百科文档中提取可验证的事件特定和参与者特定的人口统计信息,并通过自动化流程进行事实核查。DoFaiR数据集主要应用于评估和改善文本到图像模型在生成历史人物图像时的人口统计事实性,特别是在多样性干预下保持历史准确性的能力。
Demographic Factuality Representation (DoFaiR) dataset was developed by the University of California, Los Angeles (UCLA) to evaluate the balance between diversity and factuality of text-to-image models. This dataset contains 756 records covering various historical events, participants and their corresponding demographic information. The dataset creation workflow includes extracting verifiable event-specific and participant-specific demographic information from Wikipedia documents, followed by fact-checking via automated processes. The DoFaiR dataset is primarily utilized to assess and improve the demographic factuality of text-to-image models when generating images of historical figures, particularly their ability to maintain historical accuracy under diversity interventions.

- 1The Factuality Tax of Diversity-Intervened Text-to-Image Generation: Benchmark and Fact-Augmented Intervention加州大学洛杉矶分校 · 2024年



