M-Courtyard – local LLM fine-tuning app for Apple Silicon

[Bisdtsssss]

Hi MacRumors folks,

I’m the developer of M-Courtyard, an open-source macOS app for fine‑tuning LLMs
locally on Apple Silicon. I’m sharing it here to get feedback from Mac users.

It wraps the full pipeline in a GUI:
docs → training data → LoRA fine-tune → test → export to Ollama.

GitHub: https://github.com/tuwenbo0120/m-courtyard
Download: https://github.com/Mcourtyard/m-courtyard/releases/latest

Happy to answer any questions.

 

[JSRinUK]

I experimented with fine-tuning with mlx-lm a little while ago. It didn’t yield positive results, leading me to believe that the biggest hurdle in fine-tuning is curating a quality dataset - and this is something that takes a long time for your own data.

I may return to it in the future, but there’s just so much AI stuff that I want to learn right now that I don’t have a window for it. At present, I’m focussed on AI image/video generation and that has a “training” area I want to explore, too.

But I’ve bookmarked your repo, and will look back in on it when the time is right. Thanks for your efforts!

 

[Bisdtsssss]

I experimented with fine-tuning with mlx-lm a little while ago. It didn’t yield positive results, leading me to believe that the biggest hurdle in fine-tuning is curating a quality dataset - and this is something that takes a long time for your own data.

I may return to it in the future, but there’s just so much AI stuff that I want to learn right now that I don’t have a window for it. At present, I’m focussed on AI image/video generation and that has a “training” area I want to explore, too.

But I’ve bookmarked your repo, and will look back in on it when the time is right. Thanks for your efforts!

Spot on. Curating a high-quality dataset is definitely the hardest part of the process. That's actually why I built the AI-assisted data prep module into the app—to help automate the tedious task of turning raw docs into training pairs.

 

[JSRinUK]

Spot on. Curating a high-quality dataset is definitely the hardest part of the process. That's actually why I built the AI-assisted data prep module into the app—to help automate the tedious task of turning raw docs into training pairs.

My data comes from the stories I’ve written and it’s extremely difficult to automate turning those into a quality dataset (I’ve tried various methods). It’s probably more straightforward for data documents.