Nous Research released Arete-27B, a fine-tuned version of Alibaba’s Qwen3.8-27B model designed to strip out state-mandated political bias. By applying supervised fine-tuning on roughly 53 million tokens using synthetic outputs from Thinking Machines’ Inkling model, the team reduced Qwen’s state propaganda response rate from 22 percent to under 0.9 percent.
Evaluation conducted across 803 benchmark prompts scored by xAI’s Grok 4.7 showed significant performance changes on sensitive topics. On free speech queries, Arete-27B’s factual accuracy score jumped from 0.89 to 2.44 out of 5 while propaganda rates plummeted from 60 percent to 1 percent. Meanwhile, baseline reasoning and operational capabilities remained unchanged on benchmarks like Hermes Bench and Terminal-Bench 2.1.
The researchers concluded that political guardrails baked into Chinese open-weight architectures exist as a shallow outer layer. As a result, post-training adaptation without reinforcement learning is sufficient to remove regional alignment constraints without degrading core model intelligence.
Why it matters
Open-weight Chinese foundation models can be rapidly scrubbed of political censors, making them viable backbones for Western enterprise deployments.
Supervised fine-tuning (SFT) on modest token counts (50M+) is capable of overriding deeply ingrained behavioral alignment in modern LLMs.
Synthetic data distillation across international model families creates complex alignment leakage and safety evaluation challenges.
Source: the-decoder.com



