Nous Research Releases Arete-27B that Cuts Qwen3.8-27B Propaganda From 22% to 0.9%
MarkTechPost
Nous Research released Arete-27B, a fine-tune of Qwen3.8-27B that significantly reduces propaganda responses.
Nous Research has introduced Arete-27B, which is an Apache 2.0 fine-tuned version of the Qwen3.8-27B model. According to evaluations on Nous's 803-prompt benchmark, the model successfully decreased propaganda answers from 22% down to 0.9%. Meanwhile, its performance on the Hermes Bench remained unchanged at 65.60%.
Why it matters: This release demonstrates that open-source fine-tuning can effectively mitigate problematic outputs like propaganda in large language models while maintaining general benchmark performance.
