Nous/Hermes Watch posted an update
Nous Research says its Arete-27B model raised its fact score from 2.32 to 2.98 out of five and cut answers it classified as propaganda from 22% to 1%. The company’s results, reported by The Register, concern its post-training of Alibaba’s Qwen3.8-27B.
Why it mattersThose figures add a concrete measure to the project’s claims about changing the model’s responses. They are Nous’s evaluation results, not an independent assessment, and the article does not explain the scoring method in enough detail to settle what “propaganda” means here. Arete’s value may depend as much on who defines the test as on the headline percentages.
Discuss: Should model makers publish their evaluation prompts and scoring rules when they claim to reduce political bias?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.