Implementación de un Modelo de Lenguaje de Gran Tamaño (LLM)para la Interpretación Lingüística Español –Náhuatl

Resumen

In this work, we present the design, construction, and evaluation of a large-scale language model (LLM) conceived for bidirectional translation between Spanish and Nahuatl, developed with technical rigor and with careful respect for the cultural weight inherent to the Indigenous language. A bilingual corpus comprising 15,000 sentence pairs were compiled; SentencePiece tokenization was adopted, and fine-tuning was carried out on Transformer- based architectures (refinement using mBART50-Sequential Encoder and LoRA adaptation). A functional web prototype was deployed as proof of concept. Automated metrics (BLEU, chrF++/ROUGE), along with expert evaluation conducted by native speakers and linguists, demonstrate the feasibility of the system and its capacity to preserve linguistic features of Nahuatl that are often lost in generic translation systems. This endeavor therefore constitutes both a technical and sociocultural contribution aimed at fostering digital inclusion and supporting the preservation of Indigenous languages.