InterMT

该数据集名为InterMT，专为多轮次、多模态的理解和生成任务设计，通过专家标注来捕捉人类偏好。它由10万个图像-文本示例构建而成，包含了32,459个人类偏好标注，这些标注以评分形式的成对比较组织，既包括局部级别也包括全局级别。数据集规模包括15.6千个提示、52.6千个多轮对话实例以及32.4千个人类标记的偏好对。其任务重点在于多轮次多模态交互和问答。

This dataset, named InterMT, is specifically designed for multi-turn, multimodal understanding and generation tasks, and captures human preferences via expert annotations. It is constructed with 100,000 image-text examples, and contains 32,459 human preference annotations, which are organized as pairwise comparisons in the form of ratings covering both local and global levels. The dataset has a scale comprising 15.6k prompts, 52.6k multi-turn dialogue instances, and 32.4k human-annotated preference pairs. Its core task focuses on multi-turn multimodal interaction and question answering.