# Windy Translate

> Developer storefront for the open translation and speech-to-text models published by Windstorm Labs (https://windstormlabs.com), the lab of the Windstorm Institute (https://windstorminstitute.org). 1,660 translation pair models, 3 multilingual translation models and 66 speech-to-text models, covering 173 named languages. All are downloadable from Hugging Face and run locally. 788 pairs are scored on FLORES-200 dev (chrF++); the rest say why they are unscored. A hosted API is planned (early access, not live).

## Machine-readable catalogue
- [models.json](https://windytranslate.com/models.json): every model with repo, task, languages, licence, upstream base model, attribution text, NOTICE text, score (chrF++/BLEU/band) or score status. Fetch with: curl -s https://windytranslate.com/models.json | jq '.models[] | select(.task=="translation" and .src=="en" and .tgt=="fr")'
- Model pages: https://windytranslate.com/models/<id> (id = the Hugging Face repo name)

## Key pages
- [Model catalogue](https://windytranslate.com/models)
- [Models by language](https://windytranslate.com/languages)
- [Speech-to-text models](https://windytranslate.com/speech)
- [Evaluation method](https://windytranslate.com/evaluation): FLORES-200 dev, 48 sentences per pair, chrF++ and BLEU; GlotLID language verification
- [Licensing and attribution](https://windytranslate.com/licensing): Apache-2.0, CC-BY-4.0, MIT; NOTICE templates
- [Planned hosted API](https://windytranslate.com/api): not live
- [Privacy](https://windytranslate.com/privacy)

## Hugging Face
- https://huggingface.co/WindyTranslate (translation pairs, canonical)
- https://huggingface.co/WindyWord (multilingual translators and speech-to-text)

## Quick start
pip install transformers sentencepiece torch
then load any pair with AutoTokenizer / AutoModelForSeq2SeqLM, e.g. WindyTranslate/translate-tc-big-en-fr (see https://windytranslate.com/#quick-start). The transformers "translation" pipeline exists only in transformers 4.x.

Contact: hello@windytranslate.com