Nous Research releases Arete-27B, cutting a Chinese open model's propaganda rate from 22% to 1%
NousResearch · x · 2026-11-07
Nous Research post-trained Qwen3.8-27B into Arete-27B, an open model aimed at changing value alignment baked into open weights. They built an 803-prompt eval across seven value categories (free expression, democracy, privacy, etc.) measuring how often a model adopts state-aligned stances. After fine-tuning on 53M tokens, the propaganda rate fell from 22% to 1%, matching Western models tested, with no significant performance tradeoff on Hermes Bench. Eval, weights, and training data are all released, showing a model's values are one fine-tune away.
Related event: Nous Research Releases Arete-27B to Reshape Open-Model Values(2 posts)→
More from AGI Musings
- New paper 'Inter-AI Morality' studies how AI systems treat one another — jacyanthis · 2026-11-07
- Garrison Lovely: data center moratoriums won't slow AI, chip export controls would — GarrisonLovely · 2026-11-07
- New research examines the economics of AI for science: winners and bottlenecks — Afinetheorem · 2026-11-07
- AI as the new communication bridge: translating your words into the listener's mental model — narphorium · 2026-11-07
- Aryaman calls out degraded research discourse on X, says Will Depue overconfident on model internals — aryaman2020 · 2026-11-07
- New framing: AI may commoditize work rather than automate jobs away — soumitrashukla9 · 2026-11-07