The paper concerns the development of an adaptive SQL query generation system for educational purposes based on multi-stage fine-tuning of the Qwen2.5/3.0-3B language model. The investigated problem is the insufficient adaptability of existing Text-to-SQL solutions to the specific requirements of the educational process and individual requests of students. The proposed methodology combines optimized fine-tuning of the Qwen2.5/3.0-3B model with the technology of reinforcement learning from human feedback (RLHF) to create a closed loop of continuous model improvement. The system was successfully tested on a specialized dataset of 3010 examples and showed a significant improvement in accuracy: Execution Accuracy rose from 72.1% to 89.6%, Exact Match improved from 67.3% to 84.2%. At the same time, the response time decreased by 40.6%. The comparative analysis demonstrates the advantage of the proposed model over similar solutions of comparable size. The iterative improvement mechanism showed a steady increase in the quality of generated queries over three model update cycles. The obtained results have practical significance for improving the efficiency of the educational process in the field of learning SQL and business analytics based on the application of a scalable and adaptive AI solution. #COMESYSO1120