THE DEVELOPMENT AND IMPLEMENTATION OF METHODS OF OPTIMIZATION OF SMALL LANGUAGE MODELS FOR EFFICIENT GENERATION OF SQL QUERY IN EDUCATIONAL INFORMATION SYSTEMS

Author(s): Bosenko T.M.

Rubric: Information technology

DOI: 10.21777/2500-2112-2025-1-54-67

Release: 2025-1 (50)

Pages: 54-67

Keywords: small language models, optimization, SQL query generation, Text-to-SQL, quantization, pruning, educational information systems

Annotation: The article presents research on the optimization of small language models for efficient generation of SQL queries in conditions of limited computing resources of educational institutions. The objective of the study is to develop and implement optimization methods aimed at reducing the requirements for computing resources while main- taining high accuracy of SQL query generation. The methodology is based on an integrated approach, including quantization of model parameters, optimization of weight coefficients and adaptation of the dictionary to the subject area. The empirical base includes Spider, CoSQL benchmarks and a specialized DSBInf (Data Solutions Business Intelligence) dataset of 3424 queries. The developed SQL_DS_60M_ft model (a model for generating responses based on a structured query language), based on the CodeS architecture, demonstrates significant improvements: a 33 % reduction in size, a 1.4-fold increase in processing speed and an 18 % decrease in RAM usage while maintaining 65.0 % accuracy. Practical implementation of the model in the educational process showed an increase in learning efficiency by 35–40 % when working with databases and business analytics.

Bibliography: Bosenko T.M. THE DEVELOPMENT AND IMPLEMENTATION OF METHODS OF OPTIMIZATION OF SMALL LANGUAGE MODELS FOR EFFICIENT GENERATION OF SQL QUERY IN EDUCATIONAL INFORMATION SYSTEMS // Education Resources and Technologies. – 2025. – № 1 (50). – С. 54-67. doi: 10.21777/2500-2112-2025-1-54-67