Adaptive and Efficient Large Language Models for Low-Resource African Languages

Abstract

Large Language Models (LLMs) have transformed natural language processing (NLP), enabling advanced multilingual understanding and generation. Despite these advancements, African languages remain largely underrepresented due to limited datasets, high computational costs, and the scarcity of localized linguistic resources. This paper presents an adaptive and computationally efficient framework for fine-tuning open-source LLMs on low-resource African languages using Parameter-Efficient Fine-Tuning (PEFT) and synthetic data augmentation. The proposed model integrates the Low-Rank Adaptation (LoRA) mechanism into the LLaMA-3 (7B) architecture to optimize adaptation for Hausa, Yoruba, and Swahili. Training data are sourced from the AfroBench and Masakhane corpora and further enriched through back-translation and paraphrasing to enhance linguistic diversity. Experimental results demonstrate a +23% improvement in BLEU score and a 17% reduction in perplexity compared to multilingual baselines, while reducing trainable parameters by over 60%. The findings underscore the potential of efficient adaptation techniques to bridge the linguistic divide, foster inclusive AI development, and accelerate the integration of African languages into global NLP ecosystems.