Skip to content

Dataset · google

Multimodal Wikipedia image-text pairs in 108 languages

google/wit

37.6 million image-text pairs from Wikipedia across 108 languages for multimodal retrieval and training.

What's inside

The dataset is collected from Wikipedia: 37.6 million examples with entities and 11.5 million unique images across 108 languages. It includes image-text pairs and related metadata; languages include Afrikaans, Arabic, Asturian, Azerbaijani, Belarusian, Bulgarian, and others. It is suitable as a pretraining corpus for multimodal models and text-retrieval/image-to-text tasks.

What you can do with it

  • Training CLIP-like image-text models
  • Testing cross-lingual image-to-text retrieval
  • Pretraining a multimodal model on Wikipedia
Tasks
Text retrieval, Image → text
Languages
Afrikaans, Arabic, Asturian, azb, Belarusian, Bulgarian, Bangla, Breton and 4 more
Size
1M–10M rows
Downloads this month
2.8K
Likes
73

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("google/wit")