WIT Base
WIT Base is a multilingual multimodal dataset containing images extracted from Wikipedia with their associated captions in various languages, making it easy to train multilingual and multimodal AI models.
Approximately 108,000 examples (5.15 GB in the first files), multilingual images and captions, Parquet format
CC-BY-SA 4.0
Description
The dataset WIT Base is a vast collection of images with captions from Wikipedia in several languages. Each instance includes the image, its binary representation, pre-computed embedding, and a set of captions. This dataset promotes research on multimodal and multilingual models.
What is this dataset for?
- Train AI models capable of understanding and combining images and texts in several languages
- Improving multimodal translation and image description generation
- Test the robustness of models on a variety of data from the real world
Can it be enriched or improved?
Yes, it is possible to add additional linguistic annotations or to extend the dataset to other languages. Reviewing captions and standardizing can also improve quality.
🔎 In summary
🧠 Recommended for
- Multimodality researchers
- Automatic translators
- Multilingual AI developers
🔧 Compatible tools
- Hugging Face Datasets
- TensorFlow
- PyTorch
- Multimodal frameworks
💡 Tip
Use pre-calculated embeddings to speed up your experiments.
Frequently Asked Questions
What languages are covered by the WIT Base dataset?
The dataset contains data from all the languages available on Wikipedia, which ensures a very broad multilingual coverage.
Can this dataset be used in a commercial project?
Yes, licensed under CC-BY-SA 4.0, respecting the attribution and sharing clause in the same way.
Is this dataset suitable for multimodal models only?
Primarily yes, it is designed to train and evaluate models combining vision and language across multiple languages.
