Fine-tuning Large Language Models for Domain-Specific Tasks: A Comparative Study on Supervised and Parame-ter-Efficient Fine-tuning Techniques for Cybersecurity and IT Support

Sai, Chaithanya and Elmitwally, Nouh and Rice, Iain and Mahmoud, Haitham and Vickers, Ian and Schmoor, Xavier (2024) Fine-tuning Large Language Models for Domain-Specific Tasks: A Comparative Study on Supervised and Parame-ter-Efficient Fine-tuning Techniques for Cybersecurity and IT Support. In: The 4th International Conference of Advanced Computing and Informatics, 16 - 17 December 2024, Birmingham City University. (In Press)

Text ICACIN_2024_Camera_Ready_Paper_32.pdf - Accepted Version Restricted to Repository staff only Download (548kB) | Request a copy

Abstract

This study investigates the fine-tuning of open-source large language models (LLMs) for domain-specific tasks, such as question-answering in cybersecu-rity and IT support. It focuses on two fine-tuning techniques: Supervised Fi-ne-Tuning (SFT) and Parameter-Efficient Fine-Tuning (PEFT), specifically Low-Rank Adaptation (LoRA). The research compares the performance of 21 open-source LLMs, ranging from 2 to 9 billion parameters models like Llama2-7B, Llama3.1-7B, Mistral-7B, Falcon-7B, Phi-3.5, and Gemma2-9B. SFT consistently delivers high accuracy and low train-evaluation loss, while LoRA significantly reduces GPU memory usage and computational costs without compromising performance. The research findings emphasize the importance of selecting optimal fine-tuning techniques and model architec-tures for domain-specific tasks and also highlight advancements in fine-tuning LLMs for efficient and scalable AI solutions in production.

Item Type:	Conference or Workshop Item (Paper)
Dates:	Date Event 17 December 2024 Accepted
Subjects:	CAH11 - computing > CAH11-01 - computing > CAH11-01-01 - computer science
Divisions:	Faculty of Computing, Engineering and the Built Environment > College of Computing
Depositing User:	Nouh Elmitwally
Date Deposited:	13 Feb 2025 12:47
Last Modified:	13 Feb 2025 12:47
URI:	https://www.open-access.bcu.ac.uk/id/eprint/16136

Actions (login required)

View Item