Resource-Efficient Fine-Tuning of Large Language Models for Domain-Specific Text-to-SQL Tasks: A Proof-of-Concept Study with Enhanced Early Stopping and Resource Monitoring

Large language models show promise in Text-to-SQL tasks, but the high computational cost of fine-tuning presents deployment barriers in resource-constrained environments. This paper proposes a methodology combining enhanced early stopping, based on validation loss convergence, with real-time resource monitoring for efficient domain-specific fine-tuning. Experiments with Mistral 7B on the AdventureWorks database demonstrated a 40% reduction in training time while preserving model quality, with training automatically terminating after 6–8 epochs instead of the planned 12. Key contributions include a formalized early stopping framework, systematic resource-aware training, and empirical validation on an enterprise database schema.