Publication

Abstract : The research demonstrates to create an optimized system that converts natural language to SQL queries by fine-tuning Llama-3.2-1B with Spider dataset training and LoRA (Low-Rank Adaptation) for efficient parameter tuning. The experimental setup included Unsloth’s tokenizer, gradient check-pointing, and 4-bit quantization to reduce memory consumption while maintaining inference efficiency. The model achieved evaluation results through ROUGE scores (ROUGE-1: 0.1772, ROUGE-2: 0.0345, ROUGE-L: 0.1772) and cosine similarity (0.2022) to evaluate its ability to produce structured SQL queries. The CUDA acceleration and quantization techniques led to a 50% reduction in inference time which enhanced real-time usability. A Gradio UI interface enabled the system deployment to provide users with an easy-to-use interface for generating SQL queries from their input and database schema information. The system demonstrated strong performance in single-table queries yet faced difficulties with complex multi-table joins, which require future development of schema-aware learning methods. The research advances database interaction studies by developing automatic query generation systems that enable users without extensive technical skills to use SQL.