Nous Research retrained Alibaba’s open-weight Qwen3.8-27B through post-training and released Arete-27B to reinforce Western values and reduce propagandistic responses. Its factuality score rose from 2.32 to 2.98 on a five-point scale, while the proportion of responses containing propaganda fell from 22% to 1%. The team used 50 million tokens for supervised learning, with Thinking Machines Lab’s Inkling serving as the teacher model.
Source:The Register · Read original
Leave a Reply