Skip to content

Dataset · mlfoundations

DCVLM-Baseline multimodal corpus with 200B tokens

mlfoundations/dcvlm-baseline-200b

Cleaned mixture of 103.9 million image-text pairs in 6 languages in WebDataset format for VLM pre-training.

What's inside

Reference data mixture from the DataComp-VLM research paper, fully cleaned of duplicates and ready for training multimodal models. The release includes 103,985,276 examples (about 200 billion tokens) in English, Chinese, German, French, Japanese, and Korean. Data is packed into flat tar archives (WebDataset format), allowing ingestion by any training stack.

What you can do with it

  • Pre-training multimodal language models (VLM)
  • Training and evaluating visual question answering (VQA) systems in six languages
Tasks
Image text → text, Visual question answering
Languages
English, Chinese, German, French, Japanese, Korean, Latin, Finnish and 4 more
Size
10K–100K rows
Downloads this month
78.3K
Likes
9

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("mlfoundations/dcvlm-baseline-200b")