Fine-tuning open models for best-in-class Lithuanian language correction

Short description of the innovation study

Lithuanian has around three million speakers, and the large language models that dominate the market are optimised primarily for larger languages: general-purpose assistants still make frequent errors in Lithuanian grammar, spelling and punctuation. The study fine-tunes open-weight models specifically for fast, accurate Lithuanian proofreading. FOML builds a large, curated corpus of error-correction pairs from public Lithuanian sources, creates a benchmark for measuring correction quality, and uses EuroHPC resources to train and evaluate open models at a scale no SME cloud budget could support. The resulting models will run Kablelis.lt, MB Sklaida's proofreading service, at lower cost and on European infrastructure. The benchmark will be published openly, so anyone can measure how well newly released or their own models handle Lithuanian, and the study serves as a reference case for teams building language technology for other smaller European languages.

Organisations involved: 

Main SME: MB Sklaida