Parameter-Efficient Fine-Tuning with LoRA

This repository contains two separate implementations of Parameter-Efficient Fine-Tuning using Low-Rank Adaptation (LoRA), implemented by Oğuz Kağan Hitit at Koç University in 2024. LoRA allows for efficient fine-tuning of large models by adapting only a small subset of their parameters.

Running the Code

• To use the PEFT with LoRA that uses peft library, refer to PEFT_run.ipynb
• To use the custom PEFT with LoRA that does not use peft library, refer to the rest of the project.
• To run the custom implementation in Google Colab, refer to the run_in_colab.ipynb notebook, which provides step-by-step instructions and necessary configurations.

About the Implementation

LoRA's approach reduces computational overhead and preserves the original model's structure by focusing on key parameters, making it ideal for applications with limited computational resources.