TRANSFER LEARNING AND LORA ADAPTATION OF MBART-50 FOR LOW-RESOURCE BRAHUI RESPONSE GENERATION: A PERFORMANCE AND RANK ABLATION STUDY

Abstract

Languages with limited computational resources have been poorly represented in conversational AI due to the need for extensive training data and the resource requirements of neural dialog systems. Brahui, a Dravidian language spoken mainly in Balochistan, Pakistan, stands out as an especially difficult challenge because it has limited computational resources and conversational datasets. This paper examines the ability to adapt the multilingual pretrained model mBART-50 to generate Brahui conversational responses using transfer learning and parameter-efficient fine-tuning. 1,113 pairs of questions and answers in Brahui were split into 890 training samples, 111 validation samples, and 112 test samples. Full fine-tuning was contrasted with Low-Rank Adaptation (LoRA), with an analysis of ranks 4, 8, and 16. The evaluation was performed using METEOR, Translation Edit Rate (TER), and semantic similarity. Full fine-tuning scored 78% METEOR, 23% TER, and 89% semantic similarity. The LoRA with rank 8 scored the highest of all configurations tested with 80% METEOR, 20% TER, and 89% semantic similarity, whereas ranks 4 and 16 did not perform as well. It has been concluded that multilingual transfer learning may be a practical basis for developing conversational AI in Brahui, and the choice of LoRA rank apparently affects the performance of the model. The results should be considered preliminary because each configuration was evaluated in one run on a relatively small test set. The research provides an empirical basis for future work in dialogue generation in Brahui.