A hobbyist built a household chatbot that first queries a vector store, then runs each question through a classifier to narrow the search space. The classifier uses the 600‑million‑parameter Qwen 3:0.6B model, trained on roughly 850 labeled household queries ranging from pool maintenance to cooking. The goal is to map free‑form questions to a fixed metadata list before vector retrieval.

Initial prompting achieved only 10% correct classifications on a 131‑item test suite, with the model overusing generic tags and inventing out‑of‑list categories. Switching to the open‑source Unsloth framework with QLora fine‑tuning raised accuracy dramatically. After the first fine‑tune run, the model correctly labeled 104 of 131 cases, a jump to 79% accuracy.

A second iteration added a lightweight post‑processing step that normalizes partial matches (e.g., "ac" to "hvac") and replaced verbose category names with two‑character IDs to reduce semantic overlap. The experiment shows that even a sub‑billion‑parameter LLM can become a reliable, on‑device classifier when paired with a well‑curated dataset and modest prompt engineering.