Dataset · mlfoundations
MINT-1T: interleaved text and images from PDF and the web
mlfoundations/MINT-1T-PDF-CC-2023-50
Part of a massive 1 trillion token multimodal dataset for pre-training models in document and page understanding.
What's inside
The repository presents a segment of the large-scale open MINT-1T dataset from University of Washington researchers, featuring interleaved English text and images. The full corpus includes 1 trillion text tokens and 3.4 billion images, collected from sources including previously unused PDF documents and ArXiv articles; this specific split contains between 1 and 10 million records.
What you can do with it
- Pre-training multimodal LLMs (MLLMs) to work with long text-and-image context
- Training models in recognizing and analyzing complex PDF documents and research papers
- Tasks
- Image → text, Text generation
- Languages
- English
- Size
- 1M–10M rows
- Downloads this month
- 25.8K
- Likes
- 14
How to load it
from datasets import load_dataset
ds = load_dataset("mlfoundations/MINT-1T-PDF-CC-2023-50")