LLiMba adapts 3B model for Sardinian

The real story here is not just “another small LLM,” but a practical recipe for low-resource language adaptation that fits on consumer hardware.

• –The paper shows Sardinian can be meaningfully adapted with a modest GPU budget, which lowers the barrier for similar minority-language work
• –It reports stronger downstream translation performance after SFT, with rsLoRA r256 outperforming the other adapter setups tested
• –The qualitative analysis matters: some adapters look better on BLEU while still leaking scripts or fabricating more confidently
• –That makes this more useful than a vanity demo; it’s a case study in how adapter choice changes behavior, not just scores
• –The broader implication is that endangered languages may need bespoke continued-pretraining plus adapter tuning, not generic multilingual prompting