LLMs for machine translation on medium-to-low resource languages: A comprehensive evaluation for Catalan
Large language models (LLMs) that have been trained on monolingual data, predomi- nantly in English, with no intentionally included parallel text, have demonstrated remarkable potential in handling multilingual machine translation. In this study, we aim to assess the performance of decoder-only LLMs in the task of translation in medium-to-low resource languages, conducting both qualitative and quanti- tative error analyses, as well as investigating the effect of translation fine-tuning strategies to the cross-lingual transfer of the model and its ability on other tasks. All of them centred on the context of Iberian languages, focusing on the case of Catalan and Spanish.