Finetuning Workshop by Devansh Gandhi at the University of Hong Kong

This repository is to help you fine-tune Microsoft Phi 3.5 Mini Instruct to speak like you by finetuning it locally on a WhatsApp chat of yours.

Prerequisites and Requirements

Hardware Requirements

For this workshop, you'll need an Apple Silicon MacBook with at least 8GB of memory. This is necessary as we will be using MLX, a machine learning framework optimized for Apple Silicon. Other systems will not be compatible with the software we're using in this workshop.

Software Requirements

• Python 3.11
• Poetry (for dependency management)
• VSCode or Cursor or another code editor (optional but highly recommended)

Installing Python 3.11 and Poetry

1. Install Homebrew (if not already installed):

   /bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"

2. Install Python 3.11 and Poetry:

   brew install python@3.11 poetry

3. Verify the installations:

   python3.11 --version
   poetry --version

   Ensure the output shows Python 3.11.x and the version of Poetry installed.

Installation

1. Clone the repository:

   git clone https://github.com/devanshg03/finetuning-workshop.git

2. Navigate to the project directory:

   cd finetuning-workshop

3. Install dependencies with Poetry:

   poetry install

4. Activate the Poetry shell:

   poetry shell

5. Ensure the correct Python interpreter and version are selected:

   poetry env use python3.11
   poetry run python --version

   Make sure the output shows Python 3.11.x.

Preparing WhatsApp Data

1. Export your WhatsApp chat:

   • Open WhatsApp on your phone.
   • Select a direct chat (not a group chat) you want to export.
   • Tap on the contact's name at the top to open chat settings.
   • Scroll down and select Export Chat.
   • Choose Without Media to keep the file size manageable.

2. Transfer the exported chat file to your computer.

3. Place the exported chat file in the project directory:

   • Create the directory (if it doesn't exist):

     mkdir -p data/whatsapp

   • Move and rename the exported chat file:

     mv /path/to/exported/chat.txt data/whatsapp/messages.txt

     Replace /path/to/exported/chat.txt with the actual path to your exported chat file.

Note: All WhatsApp data will be stored only on your computer and will not be sent to anyone.

Data Cleaning

1. Open the data-cleaning.ipynb notebook:

   • Use Jupyter Notebook or Jupyter Lab:

     jupyter notebook data-cleaning.ipynb

2. Follow the instructions in the notebook to clean the data:

   • The notebook includes code and explanations to guide you through the data cleaning process.
   • The cleaning process is designed to accommodate different versions of WhatsApp exports. Adjust the code if necessary.

3. Generate the train, validation, and test splits:

   • Run the cells that perform data splitting to produce the datasets.

4. Save the cleaned and split data:

   • The cleaned data and splits will be saved in the data/cleaned/ directory.

Finetuning

1. Open the finetuning.ipynb notebook:

   • Use Jupyter Notebook or Jupyter Lab:

     jupyter notebook finetuning.ipynb

2. Follow the instructions to finetune your model on the cleaned data.

Additional Notes

• Code Editor: It's recommended to use a code editor like VSCode for editing scripts and notebooks. If you don't have it installed, consider downloading it.

• Visual Aids: For detailed steps with visuals, refer to the accompanying documentation and videos provided separately.

• Support for Different WhatsApp Versions: The data cleaning notebook is currently unable to handle different formats of WhatsApp chat exports. If you encounter issues, adjust the code accordingly or reach out for assistance.

• Feedback and Improvements: Your feedback is valuable. If you have suggestions or encounter any issues, please let us know so we can improve the workshop materials.