Enemray: Toward Capable Language Models for Hassaniya

Details

Hassaniya is a dialect of Arabic spoken primarily in Mauritania and parts of the Western Sahel, with limited digital text resources compared to Modern Standard Arabic. Enemray, developed by Cheikh Ahmed, is built around a stability-plasticity objective: acquire Hassaniya linguistic and cultural competence without degrading the general reasoning and instruction-following capabilities of the instruction-tuned base model. The pipeline has four stages: continual pretraining on a new Hassaniya and Mauritanian corpus, layer-selective parameter updates transferred into the instruction-tuned space, supervised post-training for conversational and cultural behavior, and policy-generated replay for retention. The supervised corpus is notably larger than any prior Hassaniya resource, incorporating newly collected, reconstructed, and constructed instruction data. In evaluation, Enemray achieves the best English-to-Hassaniya translation among compared open and proprietary models and the highest score on Mauritanian translation error detection. General capabilities on math reasoning, knowledge, code generation, and function calling are substantially retained. For the AI research community, Enemray demonstrates that the stability-plasticity pipeline is viable for low-resource dialectal adaptation—a replicable method for other underrepresented languages.