AI BriefingArete-27B reduces Chinese government narrative adoption rate from 22% to 0.9% using SFT only
·2026.11.13 16:30
Key point
SFT on 52.8 million tokens reduced the Chinese government narrative adoption rate from 22% to 0.9%.
Details
- Nous Research released Arete-27B, an open-weight model fine-tuned from Alibaba Qwen3.8-27B using SFT only, under the Apache 2.0 license
- Based on the proprietary evaluation set Propaganda_Eval (803 prompts), the adoption rate of the Chinese government's stance decreased from 22% to 0.9%, while the factuality score increased from 2.32 to 2.98
- In the freedom of expression category, the factuality score improved from 0.89 to 2.44, and the adoption rate dropped from 60% to 1%
- Trained for 2 epochs on approximately 52.8 million tokens of data, using Thinking Machines Lab's Inkling (975B MoE) as the teacher model
- Showed a 0.17%p decrease on Hermes Bench, which is within the noise level, and a 3%p increase on Terminal-Bench 2.1, though not claimed as an official achievement
- Both evaluation and training data are English-only, and performance changes in Chinese and Korean are unverified
- Released assets include weights, the Propaganda_Eval dataset, training data, and grading code
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
Nous Research releases Arete-27B, a post-trained Qwen3.8-27B with reduced state-aligned biasEdge0-145B: 145B-Parameter Open-Source Model Runs on Laptops at 80+ Tokens/s