Fine-tuning Qwen 3 0.6B: from 10% to 92% question classification accuracy

A home chatbot experiment: fine-tune a tiny local LLM with Unsloth and two-letter category codes instead of label names for metadata-aware RAG.