Skip to content

Dataset · mlfoundations

HTML portion of MINT-1T for multimodal pre-training

mlfoundations/MINT-1T-HTML

English MINT-1T HTML corpus: part of an open 1 trillion token and 3.4 billion image multimodal dataset.

What's inside

The MINT-1T-HTML repository contains HTML texts from the open multimodal dataset MINT-1T, collected by the University of Washington team. The card specifies 1 trillion text tokens and 3.4 billion images, including PDFs and arXiv articles; this is 10 times larger than existing open multimodal datasets. Language is English, size per metadata is 100M to 1B.

What you can do with it

  • Pre-training a multimodal LLM on HTML texts and images
  • Fine-tuning LoRA for image-to-text tasks
  • Evaluating models on interleaved data
Tasks
Image → text, Text generation
Languages
English
Size
100M–1B rows
Downloads this month
209.3K
Likes
133

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("mlfoundations/MINT-1T-HTML")