A Federated Adaptive Large Language Model Fine-Tuning Framework for Software Development

作者：Jianguo Chen, Zeju Cai, Wenqing Chen, Weicheng Wang, Zibin Zheng, Philip S. Yu · 发表于：IEEE Transactions on Services Computing · 年份：2025 · DOI：10.1109/tsc.2025.3623626 · 被引用次数：1 · 研究领域：Model-Driven Software Engineering Techniques、Software System Performance and Reliability、Business Process Modeling and Analysis

Large Language Models (LLMs) have achieved remarkable progress in code intelligence tasks, significantly en hancing the efficiency of software development. However, several challenges remain. First, fine-tuning LLMs for specific tasks requires a large amount of task-specific labeled data, which is often costly and time-consuming to acquire. Second, due to the sensitivity of code data, high-quality internal datasets from different organizations cannot be directly shared or combined for fine tuning. Moreover, variations in programming languages across organizations can introduce interference during the fine-tuning process. To address these challenges, we propose F-CodeLLM, a federated adaptive large language model fine-tuning framework designed for real-world software development scenarios. To the best of our knowledge, this is the first approach to apply federated learning to the fine-tuning of code LLMs, enabling collaborative model optimization while preserving the privacy of each organization's code data. We design an efficient LLM fine tuning method to mitigate the computational and communication overhead associated with collaborative fine-tuning. Experimental results demonstrate that F-CodeLLM effectively allows LLMs to learn from each organization's dataset, achieving performance comparable to centralized fine-tuning. Furthermore, F-CodeLLM is well-suited for multilingual data environments, as it can lever age shared knowledge across programming languages to enhance perf...