https://github.com/jkuri/mlx-llm-finetuning-example
LLM fine-tuning pipeline for Apple Silicon with MLX, web scraping, and LoRA training
https://github.com/jkuri/mlx-llm-finetuning-example
ai apple-silicon fine-tuning llm lora m1 m2 m3 machine-learning ministral mlx natural-language-processing
Last synced: 20 days ago
JSON representation
LLM fine-tuning pipeline for Apple Silicon with MLX, web scraping, and LoRA training
- Host: GitHub
- URL: https://github.com/jkuri/mlx-llm-finetuning-example
- Owner: jkuri
- Created: 2025-08-02T19:08:51.000Z (2 months ago)
- Default Branch: main
- Last Pushed: 2025-08-02T21:34:53.000Z (2 months ago)
- Last Synced: 2025-08-02T21:35:46.185Z (2 months ago)
- Topics: ai, apple-silicon, fine-tuning, llm, lora, m1, m2, m3, machine-learning, ministral, mlx, natural-language-processing
- Language: Python
- Homepage:
- Size: 11.7 KB
- Stars: 1
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# 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
```sh
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
```
You need huggingface account and token to download the model.
```sh
hf auth login
hf download mlx-community/Ministral-8B-Instruct-2410-4bit
```
```sh
python ./scripts/web_scraper.py https://en.wikipedia.org/wiki/Yugoslavia -p 20 -o dataset/data.csv
```
```sh
python ./scripts/prepare_jsonl_data.py dataset/data.csv
```
```sh
./scripts/train.sh
```
```sh
./scripts/test.sh
```
Example inference:
```sh
./scripts/run.sh "Explain the history of the Balkans"
```
or
```sh
./scripts/run.sh "Who was the president of Yugoslavia?"
```
## Sample