LLM Fine-tuning Pipeline on Apple Silicon GPUs with MLX

This is an LLM fine-tuning pipeline for Apple Silicon GPUs using MLX. The project enables:

Data Collection & Preparation:

Web scraping with scripts/web_scraper.py to extract content from websites (like Wikipedia) Data preprocessing with scripts/prepare_jsonl_data.py to convert scraped CSV data into JSONL training format Model Training:

Fine-tuning the mlx-community/Ministral-8B-Instruct-2410-4bit model using LoRA (Low-Rank Adaptation) Training script scripts/train.sh with configurable parameters (batch size, iterations, learning rate) Testing capabilities via scripts/test.sh Key Features:

Optimized for Apple Silicon using MLX framework LoRA fine-tuning for efficient training with limited resources Multiple data formats supported (Q&A, instruction-following, chat) Automated pipeline from web scraping to model inference

Workflow:

1. Scrape web content → CSV
2. Convert CSV → JSONL training data
3. Fine-tune model with LoRA
4. Generate responses with the adapted model

The project is designed for creating domain-specific AI assistants by training on custom web content.

Usage

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

You need huggingface account and token to download the model.

hf auth login
hf download mlx-community/Ministral-8B-Instruct-2410-4bit

python ./scripts/web_scraper.py https://en.wikipedia.org/wiki/Yugoslavia -p 20 -o dataset/data.csv

python ./scripts/prepare_jsonl_data.py dataset/data.csv

./scripts/train.sh

./scripts/test.sh

Example inference:

./scripts/run.sh "Explain the history of the Balkans"

or

./scripts/run.sh "Who was the president of Yugoslavia?"

Sample

mlx-demo.mp4