Fine-tuning Gemma using qLora and Supervised Fine-tuning

This repository contains a comprehensive notebook and tutorial on how to fine-tune the gemma-7b-it model using qLora and Supervised Fine-tuning.

Overview

This project demonstrates the steps required to fine-tune the Gemma model for tasks like code generation. We use qLora quantization to reduce memory usage and the SFTTrainer from the trl library for supervised fine-tuning.

Notebook

The notebook is available on my GitHub: gemma-Instruct-2b-Finetuning-on-alpaca.ipynb

Prerequisites

Before running the notebook, ensure you have the following:

1. GPU:

   • gemma-2b can be fine-tuned on a T4 GPU (free on Google Colab).
   • gemma-7b requires an A100 GPU.

2. Python Packages: Install the necessary packages using the following commands:

   !pip3 install -q -U bitsandbytes==0.42.0 peft==0.8.2 trl==0.7.10 accelerate==0.27.1 datasets==2.17.0 transformers==4.38.0