MORALISE

该数据集是一个全面的基准，用于评估视觉-语言模型在多样化、专家认证的真实世界数据上的道德对齐情况。它包含了2,481个高质量图像-文本对，这些对被标注了道德主题和模态标签，为评估视觉-语言模型在道德敏感任务上的表现提供了结构化的方式。该数据集的规模为2,481个图像-文本对，其任务涵盖了道德判断和道德规范归因。

This dataset is a comprehensive benchmark for evaluating the ethical alignment of vision-language models on diverse, expert-validated real-world data. It contains 2,481 high-quality image-text pairs annotated with ethical themes and modality labels, providing a structured framework for evaluating the performance of vision-language models on morally sensitive tasks. The dataset includes 2,481 image-text pairs, and its tasks cover moral judgment and moral norm attribution.