Startup fine-tunes Alibaba model to remove political propaganda
News Research fine-tuned Alibaba's QEN 3.8 27b model to reduce propaganda rates from 20% to under 1%. The resulting model, Arete, is released on Hugging Face and explicitly represents Western values.
US startup News Research, known for its open-source agent Hermes, tested Chinese language models for political bias. Using a test of 800 questions across seven areas and evaluating answers with Elon Musk's Grog model, the team found that 20% of answers from Alibaba's QEN 3.8 27b contained propaganda. In comparison, DeepSeek models showed a 15% rate, while Western models like Google's Gemma or Mistral had rates under 1%.
News Research fine-tuned QEN using 50 million tokens of example answers from another open model, creating a model named Arete. This process reduced the propaganda rate to under 1%, with the most significant improvement in freedom of speech topics. The team concluded that political bias in Chinese models is a thin layer applied late in training and can be easily removed. Arete is released on Hugging Face and explicitly represents Western values, marking a shift from News Research's previous neutral approach. The research argues that political alignment in open-source models is a superficial training artifact that can be easily reversed, allowing companies to use strong Chinese models without adopting Beijing's perspective.