Scholarlink

This paper presents a comprehensive methodology for developing and annotating specialized SQL query datasets designed for fine-tuning large language models (LLMs) for their application in personalized education systems. A five-stage algorithm for creating pedagogically-oriented datasets is proposed, including selection of educationally relevant content, development of a multidimensional annotation scheme with pedagogical metadata integration, expert annotation with inter-expert validation, iterative refinement and data finalization. Special attention is paid to principles for creating annotations that support the tasks of SQL contextual understanding, generating explanations of decision logic, automated code validation, and adaptive feedback. The methodology is validated on a case study of the sql_bi_b_db dataset (3010 records) created based on a tutorial, achieving high inter-expert agreement (κ = 0.87–0.92). The results demonstrate the critical significance of quality annotation for improving the effectiveness of LLMs in educational applications and building intelligent personalized SQL learning systems. #COMESYSO1120

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Specialized SQL Dataset Development and Annotation for Personalized LLM-Based Learning

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Albina R. Sadykova,
  • Timur M. Bosenko,
  • Polina A. Merenkova

摘要

This paper presents a comprehensive methodology for developing and annotating specialized SQL query datasets designed for fine-tuning large language models (LLMs) for their application in personalized education systems. A five-stage algorithm for creating pedagogically-oriented datasets is proposed, including selection of educationally relevant content, development of a multidimensional annotation scheme with pedagogical metadata integration, expert annotation with inter-expert validation, iterative refinement and data finalization. Special attention is paid to principles for creating annotations that support the tasks of SQL contextual understanding, generating explanations of decision logic, automated code validation, and adaptive feedback. The methodology is validated on a case study of the sql_bi_b_db dataset (3010 records) created based on a tutorial, achieving high inter-expert agreement (κ = 0.87–0.92). The results demonstrate the critical significance of quality annotation for improving the effectiveness of LLMs in educational applications and building intelligent personalized SQL learning systems. #COMESYSO1120