Conceptualization-Augmented Technical Problem Solving in TVET Education Using Fine-Tuned LLM and Polytechnic and Community Colleges Digital Dataset

Abstract

Large Language Models (LLMs) have shown strong potential for chatbot and question-answering applications, but their deployment in Technical and Vocational Education and Training (TVET) remains challenging due to domain mismatch, unreliable responses, and the need for accessible development workflows. This study aims to develop a domain-specific LLM-based chatbot for technical problem solving in TVET education by integrating fine-tuning with a PolyCC digital dataset in a cloud-based, competition-driven environment. The proposed innovation applies a structured methodology consisting of dataset preparation in instruction–response format, selection of a pre-trained LLaMA 3.2 (3B Instruct) model, parameter-efficient fine-tuning, automated evaluation using win rate, and iterative dataset optimization. The experimental findings show that model performance depends more strongly on dataset quality than on dataset size alone. From 46 experimental runs, the best-performing configuration achieved a win rate of 62%, demonstrating that carefully curated and refined question–answer data can significantly improve response relevance and domain adaptation. The study also shows that suitable epoch selection is important to balance learning and generalization, while excessive training may reduce performance. In terms of impact, this work contributes a practical framework for building domain-specific LLM chatbots in TVET, while also supporting no-code, gamified, and cloud-based AI learning for educators and students. In conclusion, the proposed approach successfully demonstrates that fine-tuned LLMs, supported by high-quality PolyCC digital datasets, can enhance technical problem solving and provide a scalable pathway for AI integration in TVET education