Latxa 7B

Latxa is a family of open large language models for Basque developed by the HiTZ Center at the University of the Basque Country (UPV/EHU), introduced March 29, 2024 (arXiv 2403.20266). The 7B variant continues pretraining Llama 2 on a new Basque corpus of 4.3M documents and 4.2B tokens — a recipe that lifted Basque performance far above every prior open model and proved competitive with GPT-4 Turbo on Basque language proficiency, while trailing on reading comprehension and knowledge-heavy tasks.

The team also released four new Basque evaluation suites (EusProficiency, EusReading, EusTrivia, EusExams) and publishes models, corpora, and datasets under open licenses. Latxa is a blueprint for low-resource language modeling, though as a 2024 research release its general capabilities lag current open models.