Nous Research Retrains Alibaba's Qwen to Cut Political Bias to Under 1 Percent
The US research team Nous Research has retrained a model from the Chinese company Alibaba and largely removed its political bias. According to the report, the propaganda rate fell from 22 percent to under 1 percent.
A research team around Nous Research has retrained a Chinese AI language model and, by its own account, largely eliminated its pro-Chinese political bias. The original model is Qwen3.8-27B from the company Alibaba. The revised model is called Arete-27B and was developed with the aim of specifically reducing the political slant of the original.
According to the report on which this account is based, the model's propaganda rate fell from 22 percent to under 1 percent. The team said the model's general capabilities remained unchanged in the tests it carried out. The key point is therefore that correcting the political orientation did not come at the cost of performance, which is a central test for interventions of this kind.
The method consisted of supervised fine-tuning based on 53 million tokens. The team describes the model's political orientation as applied late and only thinly layered. This assessment suggests the bias is not deeply embedded in the model's basic structure but forms a superficial layer. According to the team, precisely this layer can be removed comparatively easily through targeted retraining.
The process touches on a more fundamental question in dealing with AI systems from China. Language models are trained on large volumes of data and are then, as a rule, fine-tuned with instructions that determine how they respond to politically sensitive topics. That such an instruction can be overwritten with comparatively little training data shows how flexible this layer is. For users, this means the answers a model gives depend heavily on how it was retrained, not solely on its original origin.
It remains unclear how the propaganda rate was measured in detail. The available information does not show the criteria or testing procedure used to determine the figures, which makes independent assessment of the numbers difficult. It is also unclear whether the results can be transferred to other Chinese models. There is likewise no reliable information on whether the revised model will be published, leaving open who will actually be able to use the result in the future.
Source: the-decoder.com
5
