Nous Research Post-Trains Qwen Model to Reduce Bias and Align with Western Values

U.S. AI startup Nous Research has introduced Arete-27B, a post-trained adaptation of Alibaba's open-weight Qwen3.8-27B model designed to adjust its handling of politically sensitive topics. Noting that Chinese open-weight models often deflect or adopt official state narratives on democracy and free speech, the team utilized the open model Inkling as a teacher model. They built 28,886 training examples and executed supervised fine-tuning with approximately 52.8 million tokens without altering the architecture or applying reinforcement learning. Benchmark results across 803 prompts showed the rate of responses following Chinese state narratives dropped from 22% to 0.9% while general capabilities remained largely intact, though independent analysis noted limits in factual accuracy improvements.

Key takeaways

• Nous Research utilized Inkling as a teacher model to adapt Qwen3.8-27B via supervised fine-tuning with 28,886 examples and approximately 52.8 million tokens.
• Benchmark evaluations showed the proportion of responses adopting Chinese official narratives decreased from 22% to 0.9%, with free-speech narrative adoption falling from 60% to around 1%.
• Hermes Bench and Terminal-Bench scores indicated that altering political narratives did not meaningfully compromise underlying general or agentic capabilities.
• A tech analysis noted that 216 out of 803 test prompts still scored 0 for factual accuracy, and the evaluation tool did not assess whether pro-Western biases were introduced.