README.md

---
task_categories:
- zero-shot-classification
- text-retrieval
- image-to-text
- text-to-image
license: cc-by-4.0
---
This is an experimental augmentation of about 3 million synthetic captions from Recap-Datacomp-1B. This dataset includes about 2 million multilingual captions. 
It attempts to balance for gender stereotypes, added occupations, race, union membership, and religion to a subsample. We have also performed hair color and eye color balancing. It also includes some permutations of sentence orders, and modificaitons of the number of items ("Two" is changed to "Three", "Four", etc.)
We have also run the URLs of the dataset through the spawning.ai API to filter out opt-out as of June 15, 2024.

A potential use of this dataset is to generate synthetic images, including creating images of people that are more anonymized than real images.

Languages: ['bg', 'hr', 'cs', 'da', 'nl', 'en', 'et', 'fi', 'fr', 'de', 'el', 'hu', 'ga', 'it', 'lv', 'lt', 'mt', 'pl', 'pt', 'ro', 'sk', 'sl', 'es', 'sv', 'vi', 'zh-CN', 'ar', 'ru', 'hi', 'ar', 'sw', 'ja', 'ko', 'id']