OmniAlign-V
收藏资源简介:
OmniAlign-V是一个全面的多模态SFT数据集,由上海交通大学、上海人工智能实验室等联合构建,包含自然图像和 infographic图像两大类,涵盖知识性、推理性、创造性等多种任务类型,旨在通过开放的、全面的问题和回答,提升MLLMs对人类偏好的理解和响应能力。该数据集选用了丰富的语义内容图像,经过精心设计的问题和答案,形成了与现有数据集显著不同的数据分布,能够有效提升MLLMs在多模态环境下的对齐性能。
OmniAlign-V is a comprehensive multimodal supervised fine-tuning (SFT) dataset jointly constructed by Shanghai Jiao Tong University, Shanghai AI Laboratory, and other collaborating institutions. It encompasses two categories of images: natural images and infographic images, covering diverse task types including knowledge-based, reasoning, and creative tasks. The dataset aims to enhance the capability of Multimodal Large Language Models (MLLMs) to understand and respond in line with human preferences via open, comprehensive question-answer pairs. By leveraging rich semantically meaningful images and meticulously designed question-answer pairs, it creates a data distribution significantly distinct from that of existing datasets, thus effectively improving the alignment performance of MLLMs in multimodal scenarios.




