Generator Soal Cerita Kurikulum Matematika Dasar Berbasis Large Language Model

Abstract

Math word problems (MWPs) are essential for developing the reasoning skills of elementary school students; however, manual creation is time-consuming and results in limited problem variety. This study aims to design, implement, and evaluate a Large Language Model (LLM)-based math word problem generator that aligns with the Kurikulum Merdeka (Independent Curriculum) and produces valid, solvable problems in Indonesian. The system was developed using LLaMA 3.2 3B Instruct via QLoRA fine-tuning with 4-bit NormalFloat quantization, trained on a dataset of 3,185 entries covering 19 topics and 6 grade levels. The system generated 385 problems across 77 grade-topic combinations and was evaluated against a sample of 95 problems through assessments by three elementary school teachers and the G-Eval metric. Evaluation results indicate that 93 out of 95 problems met validity criteria (97.89%), with an average quality score of 4.95 out of 5.00. A comparison of the two evaluation methods revealed a Pearson correlation of 0.94 and a Cohen’s Kappa of 1.00. These findings demonstrate the system's potential as a tool to assist teachers in generating math word problems that align with the Kurikulum Merdeka.