Skip to content

Dataset · mlfoundations

Huge raw data pool for VLM training (483M)

mlfoundations/dcvlm_pool_medium

Unfiltered 41.1 TB multimodal corpus in six languages for VLM data curation experiments.

What's inside

The dataset contains 483,576,747 examples totaling 41.1 TB from 166 sources in WebDataset tar archive formats in English, Chinese, German, French, Japanese, and Korean. This is a medium-sized raw, unfiltered pool from the DataComp-VLM benchmark, not intended for direct training without preliminary cleaning. Researchers set their own filters and mixing ratios to build a custom training set based on it.

What you can do with it

  • Researching algorithms for multimodal data filtering, deduplication, and balancing
  • Building custom training sets for image-text-to-text and VQA models
Tasks
Image text → text, Visual question answering
Languages
English, Chinese, German, French, Japanese, Korean, Latin, Finnish and 4 more
Size
100M–1B rows
Downloads this month
32.8K
Likes
0

How to load it

from datasets import load_dataset

ds = load_dataset("mlfoundations/dcvlm_pool_medium")