Development of Open Large Language Models for Artificial Intelligence Digital Textbooks

Artificial Intelligence (AI) is being utilized in various fields, and research on generative AI, particularly within natural language processing (NLP) technology, is actively being conducted. Currently, research related to generative AI in the education sector utilizes closed large language models (LLMs) like GPT. However, these models have limitations, as they are difficult to fine-tune and incur high costs. This study aims to explore the potential educational applications of Open LLMs by fine-tuning and comparing the performance of Llama2 and Polyglot, which are built on a Korean-based model, with Llama3, which is not based on a Korean model. The experimental results, using a question-and-answer dataset from elementary school social studies and science subjects, showed that the Llama2 13B model exhibited the highest performance, followed by the Polyglot 12.8B model. The Llama3 8B model achieved approximately 93.08% of the performance of the Llama2 13B model and about 98.63% of the performance of the Polyglot 12.8B model. This indicates that even relatively small, non-Koreanbased models can demonstrate high performance. These results suggest that future development of LLMs base models may omit the process of converting them into language-specific base models. Additionally, fine-tuning Open LLMs for educational applications shows potential for providing personalized education.