MM-Hallu/CAST

CAST（跨模态对齐相似性测试）是一个用于视觉语言模型的数据集，包含100对从DOCCI子采样的图像。每对图像在视觉上相似但文本描述不同，旨在测试模型在文本、图像和跨模态输入上的自一致性。数据集包含图像对ID、图像、DOCCI索引、图像和描述对的相似度以及详细描述等特征。评估任务是通过不同模态识别场景相似性，使用自一致性率作为指标。

CAST (Cross-modal Alignment Similarity Test for VLMs) is a dataset containing 100 image pairs sub-sampled from DOCCI. Each pair consists of visually similar but textually distinct images, designed to test self-consistency across text-only, image-only, and cross-modal inputs in vision-language models. The dataset includes features such as pair IDs, images, DOCCI indices, image and description pair similarities, and detailed descriptions. The evaluation task involves identifying scene similarities via different modalities, using self-consistency rate as the metric.