MME-Unify
收藏资源简介:
MME-Unify是一个综合性的评估框架,旨在评估统一多模态理解和生成模型的能力。该数据集包含12个现有数据集中的任务,涵盖视觉问答、图像生成、视频理解等多种类型,共包含4104个样本。数据集的任务被分为多模态理解、多模态生成和统一任务三大类,以全面评估模型在不同模态下的理解和生成能力,以及它们如何相互增强。该数据集适用于评估统一多模态大语言模型在多模态理解和生成任务中的性能。
MME-Unify is a comprehensive evaluation framework intended to assess the capabilities of unified multimodal understanding and generation models. This dataset comprises tasks derived from 12 existing datasets, spanning a wide range of modalities including visual question answering, image generation, video understanding, and others, with a total of 4104 samples. The tasks within this dataset are categorized into three broad categories: multimodal understanding, multimodal generation, and unified tasks, enabling a comprehensive evaluation of models' understanding and generation capabilities across diverse modalities, as well as the mutual enhancement between these capabilities. This dataset is applicable for evaluating the performance of unified multimodal large language models (LLMs) on multimodal understanding and generation tasks.

- 1MME-Unify: A Comprehensive Benchmark for Unified Multimodal Understanding and Generation Models中国科学院自动化研究所, 南京大学, 北京大学, 维沃, M-M-E · 2025年



