Language verY Rare for All
- Type
- Working paper
- Status
- published
- Topics
- Machine translation, Low-resource languages
LYRA (Language verY Rare for All) combines open LLM fine-tuning, retrieval-augmented generation and transfer learning from related high-resource languages, trained on a single GPU. The study covers two-way French–Monégasque translation, a rare language unsupported by existing tools because the corpus is small. Fine-tuned open models frequently surpass and consistently match encoder-decoder systems such as NLLB on this pair.
Notes
Working paper on low-resource translation methods (LYRA), with a public copy on arXiv. Partner names stay off the related project page.