mlfoundations/MINT-1T-PDF-CC-2024-18

MINT-1T是一个开源的多模态交错数据集，包含1万亿文本标记和34亿张图像，旨在促进多模态预训练研究。数据集由华盛顿大学与Salesforce Research等机构合作创建，涵盖了HTML、PDF和ArXiv文档等多种来源。数据集通过多个步骤进行数据收集和处理，包括文档提取、过滤、图像和文本处理等。尽管数据集经过严格过滤，但仍可能存在一些偏见和风险，用户在使用时需谨慎。

MINT-1T is an open-source multimodal interleaved dataset containing 1 trillion text tokens and 3.4 billion images, a 10x scale-up from existing open-source datasets. Additionally, the dataset includes previously untapped sources such as PDFs and ArXiv papers. MINT-1T is designed to facilitate research in multimodal pretraining. The dataset was created by a team from the University of Washington in collaboration with Salesforce Research and other academic institutions including Stanford University, University of Texas at Austin, and University of California Berkeley. The creation process involves collecting documents from various sources such as HTML, PDF, and ArXiv, and performing filtering, image processing, and text processing. MINT-1T is primarily intended for research purposes and has specific guidelines for its use, including avoiding misuse and inappropriate applications. The dataset is released under a CC-BY-4.0 license.