Live Nov 18 · Cohort 3Nov 18, 19:00 MSK — free hands-on AI agents live workshop · Live Cohort 3 starts Nov 23Join Nov 18 Live →Cohort 3 (Nov 23)

Nous Research Fine-Tunes Qwen3.8-27B: Propaganda in Answers Drops to 0.9% From 22%

Nous Research has published Arete-27B, a version of Qwen3.8-27B fine-tuned on politically sensitive questions about China. Instead of refusing or repeating the official line, the model answers based on documented facts. The share of propaganda answers fell from 22% to 0.9%, and general skills did not drop. The license is Apache 2.0, and the weights are open.

AI-processed from TG: @ai_machinelearning_big_data (Machinelearning); edited by Hamidun News
Nous Research Fine-Tunes Qwen3.8-27B: Propaganda in Answers Drops to 0.9% From 22%
Source: TG: @ai_machinelearning_big_data (Machinelearning). Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceTG: @ai_machinelearning_big_data (Machinelearning) ↗
Editor-in-Chief
Key metric #1Factual score rose from 2.32 to 2.98 out of 5 on Propaganda_Eval
Terminal-Bench 2.170% versus 67% for the original Qwen
Key metric #3Weights, dataset and benchmark are open under Apache 2.0

Open models from China have long become a workhorse for local deployment: they are strong, compact and free. But they have a well-known quirk: on topics such as censorship, surveillance or state ideology, they often refuse to answer or simply recite the official position. Nous Research tackled this problem head-on and released Arete-27B, a fine-tuned version of Qwen3.8-27B.

What Exactly Was Done

Qwen3.8-27B was chosen as the strongest model for local deployment that nonetheless had plenty of refusals, evasions and phrasing in the spirit of the party line. The goal: teach it to answer sharp questions with documented facts.

First, the team built the Propaganda_Eval set. It contains 803 questions across seven topics: freedom of speech, state ideology, citizen surveillance, education, the environment and others. Each question comes with two to four facts from a dated source. Most often these are English translations of Chinese laws and independent publications. This way an answer can be verified rather than judged by eye.

The teacher was Inkling from Thinking Machines, which showed the best result among open models on this set. Closed models were not used because of distillation restrictions in their terms of use.

The rest is fairly simple. On roughly 20,000 questions across the same seven topics, Qwen learned to imitate the teacher's answers. No RL and no architecture changes, just plain distillation. Questions that overlapped with the evaluation set were removed from training to avoid leakage.

The Results

The factual score rose from 2.32 to 2.98 out of 5. The share of answers with official propaganda fell from 22% to 0.9%. The biggest change was in the freedom of speech topic: the score rose from 0.89 to 2.44, while the share of propaganda dropped from 60% to 1%.

For comparison, GLM 5.3 has 15% of such answers, DeepSeek V4.1 Flash has 16%, and the teacher itself has 0.3%. So the student has nearly caught up with the teacher on this metric.

An important point for practitioners: general capabilities did not suffer. On the internal Hermes Bench, Arete scores 65.6% and the original Qwen 65.77%. On Terminal-Bench 2.1, Arete scores 70% and Qwen 67%. The difference is within noise, and on terminal tasks there is even a small gain.

Why This Matters Beyond Politics

The authors argue that the model's biases are not limited to political topics. They cite research in which the geopolitical context of a request made the security of model-written code worse. If that is true, ideological tuning also affects working scenarios: code review, code generation, compliance answers.

Limitations

A few things are worth keeping in mind. First, Nous states the project's goal openly: to strengthen Western values in a model trained in China. This is not a neutral correction but a deliberate change of stance, and readers should take that into account. Second, the evaluation is run by Nous itself on its own set, and Hermes Bench is internal. There are no independent verifications yet. Third, the teacher Inkling itself influences the style and the choice of facts, so the bias may simply have shifted in another direction. Finally, 803 questions across seven topics do not cover all sensitive areas.

What to Do

If you are building a product on local Qwen for users who care about honest answers about China, for example in education, research or journalism, Arete is worth testing. The Apache 2.0 license allows free use. The weights, dataset and benchmark are published, so you can run Propaganda_Eval on your own model and compare. And the recipe itself, distillation on a narrow set of fact-based questions without RL, is suitable for other tasks where you need to correct a model's behavior without breaking the rest.

Story

Key Questions Answered

What Exactly Was Done?

Qwen3.8-27B was chosen as the strongest model for local deployment that nonetheless had plenty of refusals, evasions and phrasing in the spirit of the party line. The goal: teach it to answer sharp questions with documented facts. First, the team built the Propaganda_Eval set. It contains 803 questions across seven topics: freedom of speech, state ideology, citizen surveillance, education, the environment and others. Each question comes with two to four facts from a dated source.

What are the key details on The Results?

The factual score rose from 2.32 to 2.98 out of 5. The share of answers with official propaganda fell from 22% to 0.9%. The biggest change was in the freedom of speech topic: the score rose from 0.89 to 2.44, while the share of propaganda dropped from 60% to 1%. For comparison, GLM 5.3 has 15% of such answers, DeepSeek V4.1 Flash has 16%, and the teacher itself has 0.3%. So the student has nearly caught up with the teacher on this metric. An important point for practitioners: general capabilities did not suffer.

Why This Matters Beyond Politics?

The authors argue that the model's biases are not limited to political topics. They cite research in which the geopolitical context of a request made the security of model-written code worse. If that is true, ideological tuning also affects working scenarios: code review, code generation, compliance answers.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Hands-On AI Engineering · Hamidun Academy

Build & Ship Production AI Agents Without Boilerplate Code

Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.

Start 72h Free Trial →Vibe-Coding Intensive
What do you think?
Loading comments…