lballore/llimba

LLiMba (from LLM plus Limba, Sardinian for language) is an open-source project that trains an AI language model to speak Sardinian, a Romance language spoken in Sardinia, Italy that UNESCO classifies as endangered. Roughly 1 million people speak Sardinian, yet no major translation APIs or voice assistants support it. This project is a step toward changing that. The approach starts with an existing small multilingual model, 3 billion parameters, a measure of model size and capability, and adapts it for Sardinian through two rounds of training. First, the model reads around 13.9 million tokens of Sardinian text gathered from websites, Wikipedia, translated books, and poetry anthologies dating back to the 1400s. Then it is refined further on about 14,400 conversation pairs, teaching it to respond to questions and instructions in Sardinian. The resulting model can hold conversations, answer factual questions, and translate text between Sardinian and other languages. The model is designed to stay small: at its target compressed size of around 1.8 GB, it may run on a standard laptop or even a mobile device. All code, training data, and the model itself are published openly on Hugging Face under an Apache 2.0 license. A browser-based demo lets anyone try chatting with the model instantly, with no installation required. The project also compared several fine-tuning methods on the same data and hardware before settling on its final approach, and it published a technical report covering the data pipeline, training methodology, and findings. This would be useful to Sardinian language learners, researchers working on AI tools for minority or endangered languages, or developers building community services for Sardinian speakers.