MosAIC Dataset

该数据集包含了来自中国、印度和罗马尼亚的图片英文文化丰富性描述，分为三个子集：GeoDE、GD-VCR和CVQA。每位三位标注员平均为每个数据集创建了75条图片描述，特别强调了文化方面。这一规模涵盖了每个国家的图片，总共有三个数据集，任务是进行文化图像描述。This curated dataset provides culturally enriched English captions for images originating from China, India, and Romania, and is partitioned into three subsets: GeoDE, GD-VCR, and CVQA. On average, three annotators generated 75 image captions for each subset, with a particular focus on cultural dimensions. This dataset suite covers images from all three target countries, comprises three total datasets, and its core task is cultural image captioning.

搜集汇总

背景与挑战

背景概述

MosAIC是一个用于文化丰富图像描述的多智能体多模态框架数据集。该数据集专注于通过多智能体协作来生成包含文化背景的图像描述，相关研究发表于2025年NAACL会议。数据集采用MIT许可证，包含代码、模型、评估和微调等组件。