Finetuning Microsoft Phi-3 Mini 4k Instruct Model

Project Overview

This project involves fine-tuning the microsoft/Phi-3-mini-4k-instruct model using the OpenAssistant dataset. We employ the QLoRA (Quantized Low-Rank Adaptation) strategy for efficient fine-tuning. The resulting model is deployed on Hugging Face Spaces for easy access and querying.

Optimization Strategies

QLoRA Implementation

• Utilizes Quantized Low-Rank Adaptation for efficient training
• Enables fine-tuning with 4-bit quantization, significantly reducing memory usage
• Facilitates training on GPUs with limited VRAM

4-bit Quantization

• Implements 4-bit quantization using bitsandbytes
• Further optimizes GPU memory utilization, allowing for training on less powerful hardware

Project Structure

.
├── config.yaml           # Configuration for model, training, and data processing
├── finetune-phi3.py      # Main fine-tuning script
├── model_utils.py        # Utility functions for model and tokenizer setup
└── data_processing.py    # Dataset loading and preprocessing functions

Installation

Clone this repository:

git clone https://github.com/aakashvardhan/microsoft-phi3-finetuning
cd microsoft-phi3-finetuning

Install the required packages:

pip install transformers datasets peft trl bitsandbytes wandb

Set up a Weights & Biases account and log in:

wandb login

Usage

Adjust the settings in config.yaml as needed. Run the fine-tuning script:

python finetune-phi3.py

Monitor the training progress on your Weights & Biases dashboard. After training, the model will be saved in the directory specified by output_dir in the configuration.

Hugging Face Spaces

Click here to view the Hugging Face Spaces

W&B Report

Click here to view the W&B report