AI BriefingNousResearch releases Arete-27B, a Qwen3.8-27B fine-tune reducing propaganda to 0.9%
·2026.11.07 04:58
Key point
The 27B model uses supervised fine-tuning to raise fact scores from 2.32 to 2.98 across seven value categories without regressing general capabilities.
Details
- NousResearch released Arete-27B, a post-trained version of Alibaba's Qwen3.8-27B designed to reinforce Western values and remove CCP-framed answers.
- Evaluation results show the fact score increased from 2.32 to 2.98 out of 5, while answers containing propaganda dropped from 22% to 0.9%.
- The model was trained using supervised fine-tuning (SFT) only, with no reinforcement learning, preference optimization, or architectural changes.
- Improvements were observed across seven value categories, including freedom of expression, state ideology, and democracy, with no meaningful regression on general-capability or agentic benchmarks.
- The model supports a 64k context window and is intended for general assistance, agentic work, and research into value alignment.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
bf16 Fine-Tuning Freezes 95% of Weights, Masking Optimizer Instability
Find Hidden Misalignment in Training Data