mattzcarey / beginner-fine-tuning-mlx

Beginner's guide to fine-tuning with MLX (Mac only)

This repo is a beginner's guide to creating a fine-tuned text completions model. The aim is to make is as simple as possible to generate new content in the same style and tone as existing content.

Setup

Setup a virtual environment. Use any virtual environment manager.

pyenv install 3.11.7
pyenv virtualenv 3.11.7 mlx
pyenv activate mlx

Install the required packages

poetry install --no-root

Usage

Add content such as blogs or articles in md format to a blogs directory.

Run the following command to create a dataset from the content. You can optionaly push this to HF Hub with --push_to_hub flag.

python dataset/build.py

Modify the model or hyperparameters in constants.py.

Run the following command to fine-tune the model on the dataset.

python train.py

This might take a while depending on the size of the dataset and the model. You will get an estimate of the time it will take to train the model on a Mac M2 Max. If you have a Pro it will be 2x slower, a standard will be 4x of more slower. An Ultra will be 2x faster.

Evaluation

Run the following command for the vibes check.

python test.py

You can modify the prompt in test.py to see how the model performs on different prompts.

Generate

Run the following command to generate text from the model.

python generate.py --prompt "Once upon a time"

Credits to:

• mlx-examples
• Mark Lord's very helpful notebook
• unsloth