Training LoRA adapters for Apple's on-device 3B model on a free Colab T4 and a Mac

I recently posted about using Apple's on-device 3B model to build a shell assistant and benchmarking accuracy. I used the bare model and dynamic retrieval. As a next step, I wanted to explore training a LoRA adapter for the model, partly to see if it improves accuracy, but mostly to understand what Apple's training toolkit looks like and whether you can get it running on accessible hardware. Just wanted to share here in case someone is interested. Apple ships a Python training toolkit with a