Innovative Adapter-Based Fine-Tuning and Streamlined Training Strategies for Text Summarization

This research paper investigates efficient fine-tuning techniques for Large Language Models (LLMs) with a focus on minimizing computational costs and memory usage. We explore various Parameter Efficient Fine-Tuning (PEFT) approaches, including Low Rank Adaptation (LoRA), Infused Adapter by Inhibiting and Amplifying Inner Activations (IA)3, and Prompt Tuning, to evaluate their performance and feasibility in adapting open-source language models for text summarization without compromising effectiveness. By integrating DeepSpeed ZeRO-3 Offload and Flash Attention, we enhance memory utilization and accelerate training, achieving a 30-35% reduction in training time and a 15-20% decrease in GPU memory consumption. Our findings indicate that (IA)<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup> and Prompt Tuning give comparable performance to LoRA while using much lesser parameters while training. This paper provides a comprehensive analysis of PEFT techniques of different nature and their potential impact on optimizing large language models when combined with efficient training strategies, making them more accessible for consumer with less hardware resources while maintaining good performance.