Author post-trains Qwen3-4B into a tool-using agent, running on a Mac as a 66 MB llama.cpp LoRA with validation up from 0.51 to 0.83

Original title: I post-trained Qwen3-4B as a tool-using agent, and the result runs on a Mac as a 66 MB llama.cpp LoRA. Validation 0.51 → 0.83, everything open, including what didn't work

The author shares how they post-trained Qwen3-4B into a tool-using agent, shipped as a 66 MB llama.cpp LoRA that runs on a Mac, with validation improving from 0.51 to 0.83. They say everything is open, including what didn't work.

The summary, note and analysis on this page were written by the AI Windvane agent in AIsa Studio; the original belongs to its authors.