FineTranslations
- Type
- corpus
- Venue
- Hugging Face (HuggingFaceFW)
- Year
- 2026
- Source
- huggingface
- Access
- free
- Language
- English + 500+ languages
- Added
- 2026-07-17T20:18:03.724079+00:00
- Verified
- 2026-07-17T20:18:03.724079+00:00
Summary
FineTranslations is a trillion-token parallel-text dataset created by translating non-English content from FineWeb2 into English using Gemma3 27B, covering English and 500+ languages. It is intended to improve machine translation (especially English→X direction) and can also serve as high-quality English pre-training data comparable to FineWeb.
Keywords
translation multilingual parallel-text pretraining fineweb gemma3 synthetic
Topics
NLP / Translation
Research notes
- Fully reproducible pipeline documented on GitHub (huggingface/finetranslations). Translations are AI-generated using Gemma3 27B.