Skip to content

Dataset · wikimedia

Multilingual image-text pairs from Wikipedia

wikimedia/wit_base

Image and text connections from Wikipedia across many languages for training and evaluating multimodal retrieval.

What's inside

The dataset contains images and associated texts from Wikipedia articles. According to the card, it is based on WIT: over 37M image-text pairs across 108 languages; metadata specifies image and text modalities along with languages af, an, ar, arz, ast, az. The data fits image-to-text and text-retrieval tasks.

What you can do with it

  • Training a multilingual image-to-text retrieval model
  • Fine-tuning LoRA for image caption generation
  • Evaluating cross-lingual image-text search
Tasks
Image → text, Text retrieval
Languages
Afrikaans, Aragonese, Arabic, Egyptian Arabic, Asturian, Azerbaijani, azb, Bashkir and 4 more
Size
100K–1M rows
Downloads this month
11.9K
Likes
77

How to load it

from datasets import load_dataset

ds = load_dataset("wikimedia/wit_base")