Skip to content

Dataset · mlfoundations

Balanced multimodal DCVLM corpus of 200B tokens

mlfoundations/dcvlm-balanced-200b

Cleaned dataset of 112M image-text pairs in 6 languages in WebDataset format for pretraining multimodal models.

What's inside

The dataset contains 112,358,849 examples (200 billion tokens), selected and balanced from the DCVLM-large pool by the DataComp-VLM paper authors. Inside are image-text pairs in English, Chinese, German, French, Japanese, and Korean that have undergone decontamination. Data is packaged in flat WebDataset tar archives.

What you can do with it

  • Pretraining multimodal models (VLMs) from scratch on a balanced data mixture
  • Fine-tuning vision-language models for VQA and image understanding tasks across 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
65.2K
Likes
2

How to load it

from datasets import load_dataset

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