@ashen_one Just did this on a MacBook: fine-tuned Qwen3.5-0.8B (small, I know) with LoRA to call tools for my health data. The training is easy. About $3 of Claude to write the examples and a few hours of GPU (Local M4)
I fine-tuned Qwen3.5-0.8B on my MacBook (M4) to coach me on my Apple Health data.
1. Claude Haiku wrote 2,000 practice chats for $3
2. LoRA fine-tune in PyTorch
3. Broken tool calls: 31% β 2%
4. Good answers: 46% β 93%
Made a bunch of mistakes including early speed tests almost ruled out the right model because I'd forgotten one line, training crawled until I cleared the GPU memory cache every step, and my first scoring flattered the base model.
Goal is to get a bunch of tests, inject more research and data and have a recommendation like what @bryan_johnson receives from his medical team (Stretch).
I will soon write a detailed write up of the optimizations I did.
@anirudhdv27@kaggle@GoogleColab@huggingface@modal You will get 30$ in free credits and then you if you are building something, you can join their builders program (approval required) to get compute credits.