Fine-tuning Large Language Models on Justinian’s Digest for Latin-to-Italian Specialised Translation @ CLiC-it 2026

Overview

We assess the accuracy of Neural MT systems on Latin-to-Italian translation, using the parallel corpus of Justinian’s Digest, one of the most important sources of Roman legal knowledge. We evaluate several open-weight models both in zero-shot and fine-tuned settings, as well as an encoder-decoder model optimized for low-resource MT. Results show that task-specific fine-tuning yields substantial improvements across all evaluation metrics. The best-performing model, LLaMA-3.1-8B-Instruct, achieves the highest overall scores after fine-tuning. We complement state-of-the-art metrics with qualitative analysis of the generated translations. Although fine-tuned models generally produce fluent translations, challenges remain in the treatment of specialised legal concepts, terminological precision, and complex syntactic structures typical of Roman legal prose. These findings demonstrate the potential of Latin-to-Italian NMT systems as effective assistants in human-in-the-loop workflows for the translation of historical legal texts.

Models

• Hugging Face: https://huggingface.co/collections/swap-uniba/digesto-machine-translation-latin-italian

Contacts

• Eleonora Ghizztota e.ghizzota@phd.uniba.it
• Pierpaolo Basile pierpaolo.basile@uniba.it
• Lucia Siciliani lucia.siciliani@uniba.it
• Paola Marongiu paolamarongiu@cnr.it
• Alessandra Cinini alessandra.cinini@cnr.it
• Eva Sassolini eva.sassolini@cnr.it