crag-mm-2025/crag-mm-multi-turn-public

CRAG-MM是一个高质量的多模态、多轮对话基准数据集，用于评估多模态助手的视觉理解和对话能力。数据集包含关于图像的不同复杂度的对话，涵盖13个领域，包括第一人称视角图像和公共图像。数据集包括单轮和多轮两种变体，分别对应每个图像一个问题答案交换和关于同一图像的多个问题和答案的扩展对话。

CRAG-MM is a high-quality conversational benchmark for multimodal assistants, designed to evaluate the visual understanding and conversational abilities of AI systems. The dataset consists of conversations about images with varied complexity levels, covering 13 domains, including egocentric images and public images. It includes both single-turn and multi-turn variants, with single-turn being one question-answer exchange per image, and multi-turn being extended conversations with multiple questions and answers about the same image.