Arete-27B

Arete-27B is a supervised fine-tune of Alibaba's Qwen3.8-27B using 52.8M tokens over two epochs at 64k context. It was trained with no RL or architectural changes to reinforce Western values across seven categories including freedom of expression, democracy, and privacy. The model maintains near-identical general capability scores on Hermes Bench while shifting factual recall and reducing propaganda-tagged responses per the lab's LLM judge.