Nous Research introduced Arete-27B, a post-trained version of Alibaba's Qwen3.8-27B model designed to reduce Chinese state-aligned biases and reinforce Western values. Using a custom evaluation benchmark of 803 prompts, the team successfully reduced propaganda rates and improved factual recall on politically sensitive topics compared to the base model.
We post-trained Alibaba's Qwen3.8-27B to reinforce Western values on a Chinese-trained model, and measured what changed with an 803-prompt eval across seven value categories.
