Adapting Large Language Models for Arabic: Comparative Evaluation, Fine-Tuning, and Ethical Deployment

Large language models (LLMs) are reshaping information access, yet adapting them to morphologically rich and resource-scarce languages such as Arabic remains challenging. This paper presents a practical framework for Arabic-focused LLM development and deployment that integrates parameter-efficient fine-tuning (LoRA), retrieval-augmented generation (RAG), and cloud/on-premises strategies. It also introduces an ethics and observability checklist tailored to Arabic contexts. We empirically evaluate three models—TinyLlama-1.1B, Qwen2.5-7B, and DeepSeek-Qwen-14B—on Arabic medical QA and generation tasks using standard text-quality metrics (BLEU-4, ROUGE-L, and BERTScore), factuality and fluency scores, runtime measures (latency, tokens/s), and resource usage. DeepSeek-Qwen-14B achieves the highest accuracy and consistency across specialties. TinyLlama-1.1B delivers the best throughput and smallest memory footprint for edge or latency-sensitive settings, while Qwen2.5-7B provides a balanced trade-off for mid-scale deployments. RAG reduces hallucinations by approximately 12% on knowledge-intensive queries, whereas LoRA fine-tuning improves task accuracy by about 6–9% relative to RAG alone, albeit with additional training cost. We observe slight quality degradation on very long generations—especially for compact models—and variability across medical subdomains, underscoring the importance of domain coverage and dialectal diversity. Overall, the paper contributes a reproducible evaluation pipeline, model-selection guidance, and an ethical deployment checklist, highlighting hybrid RAG + fine-tuning stacks as a practical path to reliable and efficient Arabic LLMs.