MMInstruct
收藏资源简介:
MMInstruct数据集由上海人工智能实验室等机构创建,包含973K条高质量、多样化的视觉指令数据,涵盖24个日常生活中的任务领域。数据集通过结合GPT-4V和GPT-3.5以及人工校正,实现了半自动、低成本的指令生成。该数据集主要用于提升视觉大型语言模型(VLLMs)的性能,特别是在多模态指令调优方面,旨在解决现有数据集在图像多样性、标注质量和指令多样性方面的不足。
MMInstruct dataset was created by Shanghai AI Laboratory and other institutions. It contains 973K high-quality and diverse visual instruction data covering 24 daily task domains. By leveraging GPT-4V, GPT-3.5 and human post-correction, it enables semi-automatic and low-cost instruction generation. This dataset is primarily intended to enhance the performance of visual large language models (VLLMs), particularly in multimodal instruction tuning, and aims to address the shortcomings of existing datasets in terms of image diversity, annotation quality and instruction diversity.
MMInstruct
数据集概述
- 名称:MMInstruct
- 来源:论文 "MMInstruct: A High-Quality Multi-Modal Instruction Tuning Dataset with Extensive Diversity"
- 状态:即将发布
示例
- 包含一个示例图像,路径为
figs/example_in_domain.png




