Make It Work, Then Make It Better: Building a Hassaniya Machine Translation System

Hassaniya, the Arabic dialect of Mauritania, has almost no machine translation (MT) resources and no standard orthography, and it is among the weakest varieties for current large language models on recent dialectal Arabic benchmarks. A usable Hassaniya translator would let its speakers reach digital, financial, and public services in their own language. We report how we built one in two steps. We first fine-tuned NLLB-200 (1.3B) for English↔Hassaniya on about 25k sentence pairs; it reaches 59.4 chrF++ into English but only 41.5 into Hassaniya. English proves a poor pivot, since it does not mark the gender or politeness that Hassaniya requires when addressing someone. We therefore switch to Modern Standard Arabic (MSA) and fully fine-tune the Apache-2.0 model MADLAD-400 3B for Hassaniya↔MSA, with a new language tag, on four small parallel corpora (26.7k pairs after cleaning and leakage removal). To measure what each corpus contributes, we train four cumulative models and score them on a test set with an equal share from every corpus, using chrF++, spBLEU, and a copy rate that counts untranslated outputs. Zero-shot, MADLAD-400 copies 90.2% of its MSA inputs instead of producing Hassaniya. The final model reaches 48.0 chrF++ / 34.8 spBLEU for MSA→Hassaniya and 59.3 / 48.0 for Hassaniya→MSA, gains of +21.4 / +23.9 and +31.5 / +33.7 over the base model, and its copy rate falls to 6.9% (human references: 3.5%). Every corpus improves the overall score significantly, but gains are largely confined to each corpus’s own domain. A small conversational service corpus gives the largest gain after the first, and two corpora interfere in the direction that generates Hassaniya. The translator now powers a public Hassaniya chatbot that pivots through MSA to a general-purpose large language model.