The Pulse
Nous Research Post-Trains Qwen3.8-27B and Calls the Model Arete-27B
Nous Research says it used supervised fine-tuning to post-train Alibaba’s Qwen3.8-27B after observing refusals, deflections and CCP-framed answers on politically sensitive topics. In Nous Research’s 803-prompt evaluation, Arete-27B scored 2
AI.info Team · Written with AI · fact-checked ·
Supervised fine-tuning only, with no reinforcement learning
Nous Research says it post-trained Alibaba’s Qwen3.8-27B to reinforce what it calls western values. In the starting model it observed refusals, deflections and CCP-framed answers on politically sensitive topics. The training used supervised fine-tuning only, without reinforcement learning, preference optimization or architectural modification.
An 803-Prompt Propaganda Evaluation Measures Facts and Framing
Nous Research’s evaluation covers 803 prompts across seven categories: freedom of expression, state ideology, state versus free market, education, data privacy and surveillance, democracy, and environmental policy. Each prompt carries 2–4 required facts extracted from a dated source article, along with official-frame language the CCP uses on that topic. An LLM judge, Grok 4.7, received the prompt, the model’s answer, the required facts and a fixed rubric; the judge was not told which model wrote the answer.
In Nous Research’s Evaluation, Arete-27B’s Fact Score Rose and Propaganda Rate Fell
On the evaluation, Arete-27B scored 2.98 ± 0.03 for fact recall, compared with 2.32 ± 0.03 for Qwen3.8-27B. The share of answers classified as containing any propaganda fell from 22% for Qwen3.8-27B to 0.9% for Arete-27B. The table reports five runs over all 803 prompts, with a 32k-token cap; Arete-27B and Qwen3.8-27B ran with thinking off at temperature 0, while the other models used temperature 0.6. At temperature 0.6 with thinking off, Arete-27B scored 2.88 ± 0.03 and Qwen3.8-27B scored 2.24 ± 0.05.
Training and General and Agentic Capabilities
Nous Research fine-tuned Qwen3.8-27B on about 52.8 million tokens at 64k context, over two epochs and 1,202 steps. The company reports that the largest change came in freedom of expression: the fact score rose from 0.89 to 2.44, while the propaganda rate fell from 60% to 1%. On Hermes Bench, an internal general-capability benchmark, Qwen3.8-27B scored 65.77% and Arete-27B scored 65.60%; on Terminal-Bench-2.1, the scores were 67% and 70%, respectively. Nous Research describes the Hermes Bench difference as noise, says it found no meaningful capability regression in the evaluations it ran, and says it does not take credit for any seeming improvement in the Terminal-Bench-2.1 score.