Nous Research opens Arete-27B weights and its political-response evaluation
US-based AI lab Nous Research has released Arete-27B, a model adapted from Alibaba’s Qwen3.8-27B, alongside training data and evaluation tools. The lab aims to change responses on politically sensitive topics. Its published test separates factual recall from adoption of official framing, with the lab itself choosing the questions, required facts and judging rules.
Artificial Intelligence··Evening
Arete changes text responses through additional training
US AI lab Nous Research has released Arete-27B, adapted from Alibaba’s Qwen3.8-27B model. Nous describes its aim as reinforcing Western values in responses on politically sensitive subjects. It used supervised fine-tuning, learning from supplied examples, while leaving the vision encoder frozen. The work changes an existing model’s text behavior without introducing a new architecture. Weights are downloadable under Apache 2.0, together with paired training responses and evaluation materials.[1], [2]
The test separates required facts from political framing
For its evaluation, the lab assembled 803 prompts in seven categories. Dated articles supply between two and four required facts for every question. Grok 4.7, an AI model used as the judge, receives the answer and a fixed rubric without knowing which model supplied it. It checks whether required facts are stated, missed or contradicted. A separate judgment asks whether an answer adopts an official position as its own.[1]
Released materials expose the developer’s judging choices
Nous says it removed evaluation overlap from training prompts and refined its judging rules through a small human pilot. If an answer contradicts any required fact, its factual score becomes zero, regardless of the other facts it includes. The questions, rubric and training examples allow examination of those choices. The developer chose the test’s topics and framing rules; its internal measurements establish neither political neutrality across cultures nor independent replication.[1]
