State propaganda in some Chinese AI models is only "a thin layer" that fine-tuning can strip away
Summary
- Nous Research fine-tuned Alibaba's Qwen3.8-27B model using 53 million tokens of supervised fine-tuning to strip out Chinese state propaganda.
- The resulting open-weight model, Arete-27B, reduced its propaganda rate from 22 percent to under one percent without degrading general capabilities.
- Using xAI's Grok 4.7 as a judge model across 803 prompts, Nous found that political bias in tested Chinese open-weight models is superficial and removable.
- Arete-27B's training data, weights, and test dataset have been publicly released under an Apache 2.0 license on Hugging Face.
Why it matters
Demonstrating that political bias and state propaganda can be efficiently stripped from large language models via targeted fine-tuning provides a practical method for aligning open-weight models without harming general performance.