基於大型語言模型任務專屬調校：針對算術任務設計之多答案解題系統設計

Abstract

隨著大型語言模型(Large Language Model, LLM)的出現，線上教育等內容也逐漸轉型，開始使用LLM作為平台的主要核心以減少在金錢以及時間上的花費，儘管目前有許多學科已逐漸使用LLM作為知識來源，但在數學教育上的發展較為緩慢，因為數學算術問題的多變性以及解題的多樣性。在過往的研究中提出在數學學習中若給予學生多種不同的解題方向，則會提升學習者對於數學的理解以及學習的熱情。本篇研究提出運用於LLM的微調設計方法。在過往的研究中數學算術任務在NLP中一直是難以解決的領域，目前有許多研究致力於提升LLM的計算能力透過固定化計算策略，然而這樣的方式難以運用至教育領域中。本篇研究中，我們提出三種微調資料集設計之方法只需透過監督式微調技術，使得LLM能夠在一次的推理中針對數學算術問題生成多種解題方向。此微調資料集設計方法能夠在本任務中達到約77%的平均正確率。此外，本篇論文亦討論三種設計方式對多答案解系統之影響。With the emergence of Large Language Models (LLMs), online education and other content are gradually transitioning, beginning to use LLMs as the primary core of the platform to reduce costs in terms of both money and time. Although many disciplines have gradually adopted LLMs as a source of knowledge, the development in mathematical education has been slow due to the variability of its problems and the diversity of solutions. Previous studies have suggested that providing students with various problem-solving approaches in mathematical learning enhances learners' understanding of mathematics and enthusiasm for learning. This study proposes a fine-tuning design method applied to LLMs. In previous research, mathematical arithmetic tasks have always been a challenging area in Natural Language Processing (NLP). Currently, many studies are dedicated to enhancing the computational capabilities of LLMs through fixed calculation strategies; however, such methods are difficult to apply in the educational field. In this study, we propose three fine-tuning dataset design methods that only require supervised fine-tuning techniques, enabling LLMs to generate multiple problem-solving approaches for mathematical arithmetic problems in a single inference. This fine-tuning dataset design method achieves an average accuracy of approximately 77% in this task. Furthermore, this paper discusses the impact of the three design methods on multi-answer solution systems.