@article{ART003338920},
author={JongHwi Song and Urtnasan Erdenebayar},
title={Fine-tuning and Performance Evaluation of a Korean Essential Medical Domain-Specific Small Language Model Based on QLoRA},
journal={Journal of The Korea Society of Computer and Information},
issn={1598-849X},
year={2026},
volume={31},
number={5},
pages={197-205}
TY - JOUR
AU - JongHwi Song
AU - Urtnasan Erdenebayar
TI - Fine-tuning and Performance Evaluation of a Korean Essential Medical Domain-Specific Small Language Model Based on QLoRA
JO - Journal of The Korea Society of Computer and Information
PY - 2026
VL - 31
IS - 5
PB - The Korean Society Of Computer And Information
SP - 197
EP - 205
SN - 1598-849X
AB - This study proposes a method for fine-tuning a small language model (sLLM) specialized in Korean essential medical knowledge using QLoRA (Quantized Low-Rank Adaptation). We utilized the AI Hub Essential Medical Knowledge QA dataset comprising 17,280 question-answer pairs across four clinical departments—internal medicine, pediatrics, obstetrics and gynecology, and emergency medicine—to fine-tune Qwen2.5-7B-Instruct with 4-bit NF4 quantization and LoRA adapters. The fine-tuned model was trained on a single consumer-grade GPU (RTX 5070 Ti, 16GB VRAM) with only 0.92% of the total parameters being trainable. Experimental results demonstrate that the fine-tuned model achieves significant improvements over the base model: ROUGE-1 increased from 0.131 to 0.291, ROUGE-L from 0.126 to 0.291, and MCQ accuracy on the sampled internal test set (198 out of 1,728) improved from 0.0% to 68.2%, where the base model's 0.0% was primarily attributable to output format mismatch rather than lack of medical knowledge. Furthermore, on the external KorMedMCQA benchmark, the fine-tuned model achieved 58.0% accuracy compared to the base model's 55.0%. These results indicate that parameter-efficient fine-tuning with domain-specific Korean medical QA data can effectively align the output format and language consistency of general-purpose LLMs for Korean medical QA tasks, while the enhancement of medical knowledge itself remains limited and requires further investigation.
KW - QLoRA;Fine-tuning;Korean Medical QA;Small Language Model;Parameter-Efficient Fine-Tuning
DO -
UR -
ER -
JongHwi Song and Urtnasan Erdenebayar. (2026). Fine-tuning and Performance Evaluation of a Korean Essential Medical Domain-Specific Small Language Model Based on QLoRA. Journal of The Korea Society of Computer and Information, 31(5), 197-205.
JongHwi Song and Urtnasan Erdenebayar. 2026, "Fine-tuning and Performance Evaluation of a Korean Essential Medical Domain-Specific Small Language Model Based on QLoRA", Journal of The Korea Society of Computer and Information, vol.31, no.5 pp.197-205.
JongHwi Song, Urtnasan Erdenebayar "Fine-tuning and Performance Evaluation of a Korean Essential Medical Domain-Specific Small Language Model Based on QLoRA" Journal of The Korea Society of Computer and Information 31.5 pp.197-205 (2026) : 197.
JongHwi Song, Urtnasan Erdenebayar. Fine-tuning and Performance Evaluation of a Korean Essential Medical Domain-Specific Small Language Model Based on QLoRA. 2026; 31(5), 197-205.
JongHwi Song and Urtnasan Erdenebayar. "Fine-tuning and Performance Evaluation of a Korean Essential Medical Domain-Specific Small Language Model Based on QLoRA" Journal of The Korea Society of Computer and Information 31, no.5 (2026) : 197-205.
JongHwi Song; Urtnasan Erdenebayar. Fine-tuning and Performance Evaluation of a Korean Essential Medical Domain-Specific Small Language Model Based on QLoRA. Journal of The Korea Society of Computer and Information, 31(5), 197-205.
JongHwi Song; Urtnasan Erdenebayar. Fine-tuning and Performance Evaluation of a Korean Essential Medical Domain-Specific Small Language Model Based on QLoRA. Journal of The Korea Society of Computer and Information. 2026; 31(5) 197-205.
JongHwi Song, Urtnasan Erdenebayar. Fine-tuning and Performance Evaluation of a Korean Essential Medical Domain-Specific Small Language Model Based on QLoRA. 2026; 31(5), 197-205.
JongHwi Song and Urtnasan Erdenebayar. "Fine-tuning and Performance Evaluation of a Korean Essential Medical Domain-Specific Small Language Model Based on QLoRA" Journal of The Korea Society of Computer and Information 31, no.5 (2026) : 197-205.
1