Mixture of Low-Rank Adaptation Pairs of Large Language Models

by Liheng Chen

Supervised by Prof. Chuan Wu

Project Background

Large Language Models

Large Language Models (LLMs) have exhibited exceptional capabilities across various natural language processing tasks. Pre-trained on vast datasets, these models can be fine-tuned for specific tasks with remarkable efficiency, achieving state-of-the-art performances. Their deployment spans sectors including healthcare, finance, education, etc.

Parameter-Efficient Fine-Tuning

Parameter-Efficient Fine-Tuning (PEFT) methods, like P-tuning (Liu et al., 2022), Prefix-tuning (Li & Liang, 2021), and Low-Rank Adaptation (LoRA) (Hu et al., 2021), are designed to minimize the number of trainable parameters needed for model customization. These methods retain most of the model’s parameters fixed and only introduce lightweight adaptations, thus significantly reducing the memory and computation required for fine-tuning.

Project Introduction

We aim at producing a more cost-effective method to customize Large Languege Models (LLMs)

Mixture of LoRA Pairs (MoP)

  • A unified approach to optimize parameter usage during model training for downstream tasks.

Project Progress

Please stay tuned for updates!

Current stage of the project: Literature Review & Method Design

– Detailed Project Plan

Per aspera ad astra