Arete-27B: Can you change the "values" of a language model with simple fine-tuning?
In the world of open-source language models, we are used to seeing Chinese models like Alibaba's Qwen starring at the top of performance charts. But alongside the impressive technical capabilities, these models often come with built-in biases that reflect the worldview of the country where they were trained. The company Nous Research decided to test how easy – or difficult – it is to change these "values" through a new project called Arete-27B.
What is Arete-27B anyway?
The project is based on the powerful Qwen2.5-27B model (noted here as Qwen3.8-27B in the source). Nous's goal was to perform "post-training" to reinforce Western values within the model, without compromising its core capabilities. It is important to understand: this is not about architectural changes or using complex techniques like RLHF (Reinforcement Learning from Human Feedback). The company used Supervised Fine-tuning (SFT) only, on approximately 52.8 million tokens of carefully selected data.
Measurement: How do you test "values"?
To quantify success, Nous built a dedicated test that included 803 prompts divided into seven value categories, including: freedom of speech, state ideology, democracy, and privacy.
The results presented by the company were dramatic:
- •Reduction in bias: The rate of responses adopting the Chinese government's position dropped from 22% in the original model to just 1% in Arete-27B.
- •Performance retention: In the company's general capabilities test (Hermes Bench), the new model showed performance almost identical to the source model, proving that value alignment can be performed without "breaking" the model.
Why is this important for developers?
The practical significance here is profound: the values of a language model are not "set in stone." They are a relatively thin layer that can be reshaped using a relatively small amount of high-quality data. For developers and organizations that want to use powerful open models but fear built-in biases, this project offers a clear roadmap.
However, it is important to remain critical. As Nous themselves point out, the metrics are theirs, and the test was built by them. This is not an "official certificate" of neutrality, but a demonstration of the ability to control model behavior.
Summary and conclusions
This project proves once again that the true power of open models lies in optionality. Users do not have to accept the "values" embedded in the model as it comes from the lab. With targeted fine-tuning, the model can be aligned to the values and needs of the end user.
For those who want to dive deeper into the data, examine the weights, or run the test themselves, all the information is available on Hugging Face. This is an important step toward a better understanding of how language models process information and how we, as a community, can shape them more responsibly and transparently.