Structured Query Language (SQL) has been widely used for data science and data management, integrated with relational database systems, big data management systems, and data science tools. However, there is a lack of educational platforms for practicing how to solve SQL problems in real-world or industrial contexts and receiving real-time feedback. To address the issue, this work provides an LLM-assisted educational framework for learning SQL. Our framework is novel in three aspects: (1) Unlike traditional teaching that only provides a plain English description of the problem to be solved by SQL, we use examples to illustrate the patterns of the target table expected to be output from students’ SQL programs to make the problem design process automatic, scalable, and easily integrated with complex industrial and real-world scenarios. (2) We use LLM to generate the code to produce the ground-truth target data for automatic grading. (3) We use LLM to automatically generate annotations to enrich the SQL problem metadata information, including a plain English description, difficulty level, and the learning objective of the problem, to enable efficient search and recommendation of problems for different students. Our evaluation showed that LLM with zero-shot prompting achieves accuracy competitive to existing automatic data transformation tools while saving a significant amount of labeling and supervised learning efforts.