Enhancing Parameter-Efficient Fine-Tuning of Large Language Models with Alignment Adapters and LoRA

Student Name: Abdul Baseer Mohammed
Defense Date:
Location: Eaton Hall, Room 2001B
Chair: Hongyang Sun

David Johnson

Prasad Kulkarni


Large Language Models (LLMs) have become integral to natural language processing, involving initial broad pretraining on generic data followed by fine-tuning for specific tasks or domains. While advancements in Parameter Efficient Fine-Tuning (PEFT) techniques have made strides in reducing resource demands for LLM fine-tuning, they possess individual constraints. This project addresses the challenges posed by PEFT in the context of transformers architecture for sequence-to-sequence tasks, by integrating two pivotal techniques: Low-Rank Adaptation (LoRA) for computational efficiency and adaptive layers for task-specific customization. To overcome the limitations of LoRA, we introduce a simple yet effective hyper alignment adapter, that leverages a hypernetwork to generate decoder inputs based on encoder outputs, thereby serving as a crucial bridge to improve alignment between the encoder and the decoder. This fusion strikes a balance between the fine-tuning complexity and task performance, mitigating the individual drawbacks while improving the encoder-decoder alignment. As a result, we achieve more precise and contextually relevant sequence generation. The proposed solution improves the overall efficiency and effectiveness of LLMs in sequence-to-sequence tasks, leading to better alignment and more accurate output generation.

Degree: MS Project Defense (CS)
Degree Type: MS Project Defense
Degree Field: Computer Science