From the source
Nous Research post-trained
Alibaba's Qwen3.8-27B to reinforce Western values and released the result as Arete-27B, an open-weight model built on a Chinese-trained base.
- Nous says it observed refusals, deflections and CCP-framed answers on politically sensitive topics in Qwen3.8-27B and trained to alter those biases, though it does not claim to have remedied every behaviour tied to that viewpoint.
- To measure the change, Nous built an evaluation of 803 prompts across seven value categories, including freedom of expression, state ideology, democracy and data privacy, scoring answers on factual recall and on whether they adopt the Chinese state's account.
- Each prompt carries 2–4 required facts drawn from a dated source article, and an LLM judge, Grok 4.7, marks each fact stated, missed or opposed without being told which model wrote the answer.
- Nous says it chose Qwen3.8-27B because it was the most performant local model and showed much of the behaviour it targeted, and it removed training prompts overlapping the evaluation set.
- The evaluation is open, with all 803 prompts, their source articles and required facts, the judge prompt, and scripts to grade and score other models.
