Search papers, labs, and topics across Lattice.
This paper benchmarks Supervised Fine-Tuning (SFT), Low-Rank Adaptation (LoRA), and Quantised Low-Rank Adaptation (QLoRA) for fine-tuning open-source LLMs (2-9B parameters) on cybersecurity and IT support tasks. SFT achieves the best loss convergence, while LoRA and QLoRA reduce GPU memory and computational costs by 99% and 60% respectively, with only minor performance degradation. Llama, Mistral, and Phi models showed strong generalization across all fine-tuning methods.
Achieve comparable performance with 99% less GPU memory by using LoRA and QLoRA for fine-tuning LLMs on cybersecurity and IT support tasks.
The focus of Artificial Intelligence (AI) has progressed from exploring whether Large Language Models (LLMs) can transform domain-specific tasks to understanding how to optimise their impact. Advances in Natural Language Processing (NLP) have enabled LLMs to address complex tasks such as question answering, text classification, and translation with notable accuracy and efficiency. Hence, this study evaluates the effectiveness of fine-tuning open-source LLMs, which range from 2 to 9 billion parameters, for cybersecurity and IT support using three advanced techniques: Supervised Fine-Tuning (SFT), Low-Rank Adaptation (LoRA), and Quantised Low-Rank Adaptation (QLoRA). The results indicate that SFT achieves superior convergence in loss metrics for both tasks, while LoRA and QLoRA significantly reduce GPU memory and computational costs by 99% and 60%, respectively, with comparable performance. Models such as Llama, Mistral, and Phi consistently demonstrated high generalisation across all methods, making them particularly suitable for deployment in resource-constrained environments.
