Recent advancements in Natural Language Processing (NLP) and Text-to-SQL systems have enabled easier interaction with relational databases. However, most solutions focus on English, leaving Arabic underrepresented. This study addresses that gap by fine-tuning the Llama-3-SQLCoder model to convert Arabic text into correct and executable SQL, enabling non-technical users to work with databases without learning SQL syntax. We enhanced the model using Low-Rank Adaptation (LoRA) and Unsloth, training it on Arabic questions paired with SQL queries from the Northwind database. To support low-resource environments, the model was converted to the GGUF format, reducing computational requirements while preserving performance. Evaluation results showed an execution accuracy of 90.24% and a validity rate of 97.56%, outperforming the zero-shot baseline (44% and 80%). The model also achieved an Exact Match score of 32% and an F1 score of 0.83, compared to 12% and 0.61 for the baseline. These findings demonstrate that LoRA and Unsloth are effective for adapting SQL-specialized models to Arabic. Despite these improvements, the system still struggles with complex nested queries and dialectal variations, indicating areas for future work. Overall, this study contributes to narrowing the gap between Arabic and other languages in Text-to-SQL research and improves database accessibility for non-technical users.
