A personal project demonstrates that a tiny 0.6B-parameter local AI model can be fine-tuned to reliably categorize household questions, jumping from 10% to 79% accuracy.

What happened: A developer fine-tuned Qwen 3:0.6B—a very small language model with 600M parameters—to categorize household questions into predefined categories like pool, hvac, and cooking. Using the Unsloth framework and a dataset of roughly 850 labeled questions, the model's accuracy improved from ~10% (baseline, prompt-only) to 79% on a held-out test set of 131 questions.

Why it matters: The baseline model failed badly, confusing categories and inventing ones not on the allowed list. After fine-tuning on household-specific data, the model now correctly routes most questions to the right metadata bucket, enabling a downstream vector database search to return more relevant household knowledge. This suggests that even tiny local models can become useful classifiers when trained on task-specific data.

What to watch: The fine-tuned model still makes errors—it sometimes returns fragments of category names (e.g., 'ac' instead of 'hvac') and struggles with semantically overlapping categories like water heater, fountain, and pool. The developer is exploring further refinements to the training approach, rather than post-processing patches, to improve handling of these edge cases.