Tao: Using test-time compute to train efficient LLMs without labeled data