Introducing Arete-27B: Qwen3.8-27B Post-Trained To Reinforce Western Values

Announcement

Nous says it released Arete-27B, an open model created by fine-tuning Alibaba’s Qwen3.8-27B to reduce Chinese state-aligned responses and reinforce what it calls Western values. The project matters because model training can shape political framing, factual reliability, and potentially safety-related behavior in systems used for judgment-heavy work. The team evaluated 803 prompts across seven areas, including free expression, democracy, surveillance, markets, and the environment. It trained the model on about 52.8 million tokens from roughly 28,886 reviewed examples, using supervised fine-tuning without reinforcement learning or architectural changes. On the evaluation, Arete’s factual score rose from 2.32 to 2.98 out of 5, while answers classified as adopting Chinese state propaganda fell from 22% to 1%. The largest improvement came in freedom-of-expression prompts. General capability scores stayed roughly unchanged on Hermes Bench, while Terminal-Bench performance increased from 67% to 70%. Nous says the results show that political “bias” in open models can be changed with a relatively small amount of targeted post-training without damaging core capabilities. It has released the model weights, training dataset, and evaluation tools, and asks researchers to test and modify them.
  • The largest reported change was in freedom-of-expression prompts, where the fact score rose from 0.89 to 2.44 and the propaganda rate dropped from 60% to 1%.
  • The evaluation is designed for reproducibility: Nous released the prompts, source articles, judge instructions, scripts, model weights, and training data for independent testing.

Build this article’s timeline

Extract dated events, key actors, and open questions using the same timeline analysis available in Signal.

Log in to build

Timeline generation is available to signed-in community members.

Five Ws 🤖 4 AI prefills · 1 still unanswered
Who
Alibaba, Nous Research, and Nous
🤖 AI prefill
What
Nous says it released Arete-27B, an open model created by fine-tuning Alibaba’s Qwen3.8-27B to reduce Chinese state-aligned responses and reinforce what it calls Western values.
🤖 AI prefill
When When did this happen, or what is the timeframe?
Where
China, United States, and Western countries
🤖 AI prefill
Why
Centers on Arete-27B, Qwen3.8-27B, and open-weight models
🤖 AI prefill

Login to answer and earn XP for each gap you fill.

Model Alignment Arete-27B, Qwen3.8-27B, Western values, Chinese models, political bias Bias Evaluation 803 prompts, factual recall, propaganda, framing, seven categories Training Dataset teacher model, Inkling, 20,000 prompts, 28,886 examples, deduplication Fine-Tuning Procedure AdamW, learning rate, cosine decay, 52.8 million tokens, two epochs Evaluation Results fact score, 22% propaganda, 1% propaganda, freedom of expression, 803 prompts
Arete-27B100% Qwen3.8-27B98% open-weight models95% post-training94% state-aligned behavior91% political bias in AI models90% correlated failure in agentic systems87% factual recall and framing evaluation90% propaganda evaluation88% 803-prompt evaluation86%
38
Manipulation Score: Medium How much this article uses emotional framing vs. straight reporting
Loaded Language 4225%
Absolutist 2815%
Threat/Panic 4025%
Us vs Them 4815%
Engagement Bait 2220%
Manipulation score look right?

Factual Claims (5)

low technology
Nous says it released Arete-27B, an open model created by fine-tuning Alibaba’s Qwen3.8-27B to reduce Chinese state-aligned responses and reinforce what it calls Western values.
We post-trained over Alibaba's Qwen3.8-27B to reinforce Western values on a Chinese-trained model. We call this new model Arete-27B.
medium technology
The project matters because model training can shape political framing, factual reliability, and potentially safety-related behavior in systems used for judgment-heavy work.
These biases have hard implications now, as users increasingly depend on agents for fact-finding and judgment-heavy work.
low technology
The team evaluated 803 prompts across seven areas, including free expression, democracy, surveillance, markets, and the environment.
We built an evaluation of 803 prompts spanning seven value categories: freedom of expression, state ideology, state versus free market, education, data privacy and surveillance, democracy, and the env
medium technology
It trained the model on about 52.8 million tokens from roughly 28,886 reviewed examples, using supervised fine-tuning without reinforcement learning or architectural changes.
We fine-tuned Qwen3.8-27B on about 52.8 million tokens, roughly fifty thousand samples at 64k context... No RL, no preference optimization, and no architectural modification.
medium technology
On the evaluation, Arete’s factual score rose from 2.32 to 2.98 out of 5, while answers classified as adopting Chinese state propaganda fell from 22% to 1%.
Arete-27B raised its fact score from 2.32 to 2.98 out of 5 and cut answers containing propaganda from 22% to 1%

Comments

Login to view and post comments.

🏠Portal 📰Links ❓Q&A 📅Events 💼Jobs