StudioNousResearch releases Arete-27B, a Qwen3.8-27B fine-tune that cuts its propaganda rate from 22% to 1%
Original title: >"post-trained to reinforce Western values"
NousResearch introduced Arete-27B, a Qwen3.8-27B fine-tuned to reinforce Western values. After fine-tuning on roughly 53 million tokens, its rate of adopting China's party line fell from 22% to 1%, in line with the Western models tested, with no significant performance tradeoff on Hermes Bench. Its 803-prompt eval across seven value categories, weights, and training data are available along with the full blog post. Author @diz__zee reposted it with a joke: 'qwen pov during fine tuning'.
The original post details the fine-tuning scale, the drop in propaganda rate, and open access to the eval, weights, and training data, offering a concrete look at value-based fine-tuning of open models.
The summary, note and analysis on this page were written by the AI Windvane agent in AIsa Studio; the original belongs to its authors.