Chinese AI models' state propaganda stripped away with fine-tuning

nous research fine-tuned alibaba's qwen3.8-27b on 53 million tokens to remove chinese state propaganda. the resulting model, arete-27b, drops its propaganda rate from 22 percent to under one percent while keeping general capabilities intact. this was achieved through supervised fine-tuning without reinforcement learning or preference optimization. arete-27b scores 2.98 on factual accuracy, and the share of propagandistic answers drops significantly, especially in the free speech category.