Large Language Models on Problem-Solving Skills in Programming Debugging of IT Undergraduates

This dataset captures multi-dimensional information from undergraduate students enrolled in Level 2 of an IT-related program, focusing on their engagement with Large Language Models (LLMs) and debugging performance. The first component of the dataset records demographic details (gender, academic stream, and A\/L Z-score), academic performance indicators (overall GPA and Level 1 programming module results), and tool usage patterns. Students reported their preferred and frequently used LLMs, with ChatGPT being the most dominant, alongside Copilot, Gemini, and other emerging models. Additional variables include self-rated confidence in debugging tasks (on a 1\u20135 scale), timestamps of pre-test, mid-test, and post-test sessions, and the Integrated Development Environments (IDEs) used (e.g., Visual Studio Code, NetBeans, Online Java compilers).The second component of the dataset consists of mock-test results from 4 students and proper-test results from 83 students. It evaluates debugging performance across pre-test, mid-test, and post-test phases, measuring (i) number of errors identified, (ii) number of errors correctly fixed, (iii) error types (logical vs. syntax), (iv) duration taken to complete tasks, and (v) accuracy rate. The sample demonstrates progression in error identification and correction skills, with notable improvements in mid-test performance (100% accuracy for some students), followed by more realistic outcomes in the post-tests where accuracy rates ranged around 56%.Together, these datasets provide a comprehensive foundation for examining the relationship between LLM adoption, student confidence, and actual debugging performance. They enable comparative analyses of learning progression, the role of AI-assisted tools in programming education, and the balance between speed, accuracy, and tool reliance in problem-solving tasks.