Research team points out Chinese AI model propaganda is a 'thin layer' removable via fine-tuning
Key point
Nous Research retrained Alibaba's Qwen3.8-27B to create Arete-27B, reducing pro-Chinese bias from 22% to less than 1%.
Nous Research announced the development of Arete-27B by retraining Alibaba's Qwen3.8-27B. This process utilized 53 million tokens of supervised fine-tuning.
According to the team, the propaganda rate in the original model was 22%, but it dropped to less than 1% in Arete-27B. They stated that general capabilities remained intact in tests. The team writes that political bias is "applied late, it is applied thin."
This item is a summary. The full, checkable version is at the original source:
Read the original at The Decoder ↗