Skip to content

Dataset · HuggingFaceFW

Parallel texts in 500+ languages

HuggingFaceFW/finetranslations

Over 1 trillion tokens of parallel text in English and 500+ languages for machine translation.

What's inside

Parallel texts in English and 500+ languages, over 1 trillion tokens. Generated by translating FineWeb2 data into English using Gemma3 27B. The synthetic pipeline is deployed via datatrove with inference managed on VLLM.

What you can do with it

  • Train machine translation models.
  • Fine-tune multilingual models for low-resource languages.
  • Evaluate translation across 500+ languages.
Tasks
Text generation, Translation
Languages
Abkhazian, abq, abs, acm, adh, adi, Adyghe, Tunisian Arabic and 4 more
Size
1B–10B rows
Downloads this month
27.3K
Likes
310

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("HuggingFaceFW/finetranslations")