Good results fine tuning a local LLM like Qwen 3:0.6B to categorize questions