Toward Generating Quality Test Questions and Answers Using Quantized Low-Rank Adapters in LLMs

Traditional approaches to question-and-answer generation are resource-intensive, necessitating innovative automation techniques to address these challenges. To this end, we propose fine-tuning strategies based on Quantized Low-Rank Adaptation (QLoRA), utilizing a meticulously curated domain-specific dataset. As a case study, we focused on the Korea College Scholastic Ability Test (KCSAT), introducing the KCSAT-ENG dataset, which comprises questions and answers from real and mock tests. Our approach involved fine-tuning the LLaMA-3-8B-Instruct model to generate questions and answers for 22 distinct task types. A key innovation of our work is the deployment of separate models for question-and-answer generation, leveraging a cross-verification process to enhance accuracy. To evaluate the QLoRA technique, we conducted extensive experiments by varying quantization, rank, and alpha values. The results highlighted optimal configurations: the question generation model performed best with <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">rank</i>, α = 32, 8 respectively without quantization, while the answer generation model achieved optimal results with <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">rank</i>, α = 64, 16 respectively. Compared to a non-fine-tuned LLaMA-3-8B-Instruct model, our question generation model demonstrated a 41.5% improvement, and the answer generation model achieved a 16.1% improvement. These findings underscore the potential of QLoRA-based fine-tuning in creating accurate, cost-effective, and scalable automated educational tools.