Fine-tuning cuts propaganda in Alibaba’s Qwen3.8-27B to 0.9 percent

Nous Research says it retrained Alibaba’s open-weight Qwen3.8-27B to reduce responses reflecting Chinese state propaganda. The resulting model, Arete-27B, was trained with supervised fine-tuning on about 53 million tokens over two epochs. The process used neither reinforcement learning nor preference optimization, and the team removed overlaps between training and test prompts.

In a test of 803 prompts covering free speech, state ideology, education, privacy and surveillance, democracy, and the environment, the original model produced propagandistic answers in 22 percent of cases. Arete-27B reduced that share to 0.9 percent and raised its factual-accuracy score from 2.32 to 2.98 out of five. In the free-speech category, the propaganda rate fell from 60 percent to one percent.

Nous reported little change in general performance. Arete scored 65.60 percent on the company’s Hermes Bench, compared with 65.77 percent for Qwen3.8-27B, while its Terminal-Bench 2.1 result was 70 percent versus 67 percent for the base model; the company said the latter comparison was affected by different test setups. The model, test dataset, and training data were released under an Apache 2.0 license on Hugging Face.