22% to 0.9%: Nous Research Tames the Propagandistic Tone of Qwen3.8-27B

Nous Research has released Arete-27B, an Apache 2.0 licensed fine-tuned model derived from Qwen3.8-27B. MarkTechPost author Asif Razzaq shared a very specific comparison result: in Nous's internal 803-prompt evaluation, the share of propagandistic answers dropped from 22% to 0.9%, while the Hermes Bench score showed no obvious fluctuations, remaining at 65.60%.

First let's add the missing context. The "propagandistic answers" referenced here refer to the percentage of cases where the model outputs propagandistic, stance-over-fact content when responding to targeted questions. The 803 prompts come from internal evaluation, which is still far from covering all real-world use scenarios. The 0.9% figure also merits ongoing monitoring: it shows this type of response has been suppressed to a very low level, but the model could still display the same tendency when given other prompts.

What stands out is that the Hermes Bench score stayed flat. For behavioral fine-tuning, a common risk is that you fix one flaw but accidentally erode the model's general capabilities in the process. The public figures at least confirm that on this set of assessments, the de-propagandization adjustment did not cause any noticeable capability regression. As for training data, methodological details and performance across a wider range of tasks, the original article did not expand on these points, so full reproduction and proper judgment will require access to the model weights and follow-up testing.

My take: The value of this type of fine-tuning lies in deployment. When large language models are integrated into customer service, content generation, and education products, what users notice first rarely relates to parameter scale — it almost always comes down to how the model communicates. De-propagandization sounds like value alignment engineering, but in practice it is more about trimming the model's output distribution. A single benchmark cannot prove the model is fully neutral permanently, but the result of cutting the propagandistic response rate from 22% to 0.9% with no drop in the Hermes Bench score is clear enough on its own.

Original article link: https://www.marktechpost.com/2026/11/07/nous-research-arete-27b-qwen3-8-27b-propaganda-fine-tune/