TIGER-Lab/VisualWebInstruct

VisualWebInstruct是一个大规模多样化的多模态指令数据集，包含约90万个问题-答案对，旨在提高视觉语言模型的推理能力。数据集覆盖了数学、物理、金融、化学、工程等多个学科，经过精心策划和筛选，强调多步骤推理任务而非简单的基于感知的问题。

VisualWebInstruct is a large-scale, diverse multimodal instruction dataset containing approximately 900K question-answer pairs, designed to enhance the reasoning capabilities of vision-language models. The dataset covers multiple disciplines including Mathematics, Physics, Finance, Chemistry, Engineering, and others, and is carefully curated and emphasizes multi-step reasoning tasks rather than simple perception-based questions.