Tom9358/nllb-gos-nld-v2

Model Card for Model ID

Moi!

This is a finetuned NLLB model that adds support for Gronings (gos). It was trained on sentence pairs from Tatoeba and on synthetic data generated from the Gronings-language Kreuze magazine. Machine translation is never perfect, and this project is also a work in progress. In my experience the model can produce reasonable Gronings from Dutch input a lot of the time, but it should not be trusted blindly. Consider this a beta release!

Model Details

The base model is a 1.3B NLLB translation model (hat was distilled from a much larger model. I finetuned using Gronings data from Tatoeba supplemented with synthetically generated data from the Gronings Kreuze magazine. I finetuned the model on multilingual Tatoeba data to avoid critical forgetting of its multilingual translation capabilities. Training is done in two stages: first a stage with synthetic data alongsize tatoeba data, and then a clean finish of only non-synthetic data from Tatoeba. BLEU scores and ChrF scores can be found on github. In addition to this I evaluated by manual inspection, as I am a linguist and speaker of Gronings. The translations are in my opinion good enough that the model can be published here and be tried out by others, and I'm hoping that in the near future it can be put into practice in the real world as well.

Model Description

• Developed by: Tom Brand (me)
• Funded by: Nobody. I did this in my free time.
• Shared by: Me
• Languages used (Tatoeba training data): Gronings, Dutch. YMMV with German, Spanish, and French.
• Finetuned from model: NLLB 1.3B distilled

• Repository: https://github.com/tom9358/nllb-tryout
• Demo: https://huggingface.co/spaces/Tom9358/gos_gronings_translate

For more documentation, information about the sources, and things I've tried, see the repo and its issues.

Don't doubt to contact me for anything related to this project! The easiest way is probably via a github issue