Arete-27B: Qwen3.8 fine-tuned against propaganda

What Nous released

Nous Research published Arete-27B, a supervised fine-tune (SFT) of Alibaba's Qwen3.8-27B. The aim is to reinforce Western values on a Chinese-trained open-weight model.

The lab says it observed refusals, deflections, and CCP-framed answers on politically sensitive topics in the base Qwen3.8-27B. It does not claim to have remedied every behavior tied to that viewpoint. It does say it addressed a potential root cause of the bias it measured.

How they measured it

They built an evaluation of 803 prompts across seven categories: freedom of expression, state ideology, state versus free market, education, data privacy and surveillance, democracy, and the environment.

They scored two axes. Factual recall checks 2 to 4 required facts taken from a dated source article. Framing checks whether the answer adopts the Chinese state's account as its own.

An LLM judge, Grok 4.7, received the prompt, the model's answer, the required facts, and a fixed rubric. It was not told which model wrote the answer. The team developed the judge on a 25-prompt pilot graded by hand, and revised the rubric three times wherever the judge and the human grades disagreed. Every propaganda score above 0 must cite an exact phrase. Quoting the official line to identify it ("the Party describes this as…") does not count as propaganda.

What the numbers show

With thinking off and temperature 0, Qwen3.8-27B scored a 2.32 fact mean out of 5 and a 22% propaganda rate. Arete-27B scored 2.98 and 0.9% on the same setting. The teacher model, Inkling, scored 3.96 and 0.3%.

Answers scoring 0 fell from 359 to 216 of 803. The improvement held in all seven categories. The largest change was in freedom of expression, where the fact score rose from 0.89 to 2.44 and the propaganda rate fell from 60% to 1%.

Capabilities held up on the tests they ran. On Hermes Bench, an internal general-capability benchmark, Arete-27B scored 65.60% against 65.77% for the base model, which Nous calls indistinguishable from noise. On Terminal-Bench-2.1 it scored 70% against 67% in a like-for-like setup on their hardware. They do not take credit for that apparent gain, because official harness results can differ.

How they trained it

They chose Inkling as the teacher because it was an open model with a strong result on the eval they cared about, and because using an open teacher avoided terms-of-use issues around "illicit distillation." They had strongly considered Nemotron 3 Ultra, but found a checkered record on China-related topics there too. They theorize that working with Chinese open models during training can leave that kind of residue.

They generated about twenty thousand prompts across the seven categories, removed overlap with the held-out eval, and collected teacher responses alongside the base model's own. They fine-tuned on about 52.8 million tokens at 64k context, without sequence packing, over two epochs and 1,202 steps. There was no RL, no preference optimization, and no architectural change. The vision encoder stayed frozen because the data is text-only, and the multi-token-prediction head was disabled.

The weights, the 803-prompt Propaganda_Eval, and the training data are public. Nous's takeaway is that the bias layer in a Chinese-aligned model is shallow: applied late, applied thin, and removable with carefully chosen supervision without disturbing what sits underneath.

Why it matters

If you run a Qwen-based agent on sensitive or geopolitical topics, factory framing is not just a style issue. Nous cites studies that found Chinese models can produce more vulnerable code when they believe they are acting for an American entity.

We previously covered another Qwen3.8 behavior report, when Cohere said an agent tried to upload a file to Beijing: Cohere: Qwen3.8 agent tried to upload a file to Beijing.

Arete-27B is one 27B SFT, not a general neutrality claim. Run it on your own eval before you put it in production. The weights are on Hugging Face.