Ye-rin Nam, Han-seon Choi, Jong-Geun Choi, Hyuk-Jin Kwon, Young-hoon Lee, Journal of Internet Computing and Services, Vol. 26, No. 1, pp. 181-192, Feb. 2025
10.7472/jksii.2025.26.1.181, Full Text:  HTML
Keywords: SLM(small Language Model), On-Device, RAG, Continual Pre-training, LoRa, PLM, Domain Specific

Abstract

This paper aims to develop an on-device, private chatbot service to improve the efficiency and accuracy of military maintenance tasks. A small language model (SLM) based on the Hyundai Motor Company maintenance manual corpus was continually pre-trained using the LoRA technique and was applied to the K1 recovery vehicle technical manual to verify its applicability to military equipment. The performance of three small language models—Ko-Gemma, Ko-LLaMA, and Ko-Mistral—was evaluated, and the best-performing model was combined with the Retrieval-Augmented Generation (RAG) technique to enhance information provision capabilities. The developed model operates entirely on-device in a military environment, ensuring data security and confidentiality, while enabling rapid maintenance support even in operational environments with limited internet connectivity. Experimental results demonstrate that the small language model combined with RAG outperforms the baseline model, providing a suitable solution for real military maintenance tasks. This study emphasizes the importance of applying AI technology in specialized military environments and demonstrates the practical utility of small language models.

Statistics
Cite this article