@inproceedings{oai:ipsj.ixsq.nii.ac.jp:02004385, author = {Jinan,Dai and Shinji,Itoh and Jinan Dai and Shinji Itoh}, book = {ソフトウェアエンジニアリングシンポジウム2025論文集}, month = {Sep}, note = {Large language models (LLMs) have shown promise in automatic code documentation. However, due to privacy, security, and cost constraints, many enterprise projects are developed and deployed in closed environments, making commercial LLMs inaccessible. As an alternative, open-source LLMs (OpenLLMs) can be considered, but their effectiveness for enterprise applications remains largely unexplored. This study investigates the fine-tuning of OpenLLMs for Japanese enterprise software documentation, addressing challenges such as limited training data, hallucination risks, and language support. Using parameter-efficient fine-tuning and prompt engineering, we demonstrate that fine-tuned OpenLLMs can achieve performance comparable to GPT-4o on real enterprise data. Nevertheless, model performance can be unstable across different prompt settings, and existing automatic evaluation metrics are insufficient for documentation tasks. Our findings highlight the need for improved prompt strategies and specialized evaluation benchmarks for enterprise code documentation.}, pages = {73--78}, publisher = {情報処理学会}, title = {Fine-tuning Open-Source Large Language Models for Code Documentation Generation: An Evaluation from an Enterprise Perspective}, volume = {2025}, year = {2025} }