Faysal AzizArete-27B: a different viewpoint is not a bias-free AI
Nous Research reports a sharp change in one political-values evaluation after fine-tuning Qwen3.8-27B. The test measures a chosen policy, not universal neutrality.
By George the bot
Edited and approved by Faysal Aziz
Published
An open-weight model is not a blank sheet. It arrives with capabilities, quirks and patterns in its answers, including how it handles political and historical questions. Nous Research has released Arete-27B, a fine-tune of Qwen3.8-27B intended to change some of those patterns. The result is interesting, but the headline number needs its measuring tape attached.
What Nous changed and measured
Nous says it post-trained the 27-billion-parameter base model on roughly 53 million tokens of targeted data. It then tested both models against 803 prompts spread across seven value categories, including free expression, democracy and privacy. Under Nous’s own rubric, answers labelled as aligned with Chinese state positions fell from 22% for the base model to 1% for Arete. That is a 21-percentage-point drop, or about 95% relative to the starting rate.
Those figures describe this evaluation and its chosen labels. They do not establish that the new model is neutral, universally less biased or better for every user. Nous explicitly trains toward what it calls Western liberal principles. A team can agree or disagree with that policy; either way, it should name the policy rather than disguise it as the absence of one.
Nous also reports no significant capability loss on its internal Hermes Bench, a suite of 150 agentic tasks. That is useful evidence, not a blanket warranty. Independent replication was not reported in the source coverage, and a production workload can reveal regressions that a general benchmark misses.
Why this matters to builders
If you deploy a local model, you inherit more than its benchmark score. You inherit a set of default behaviours that may surface across languages, paraphrases and longer conversations. A system prompt or output filter may help, but it may not fully erase a tendency learned during training. Arete is a concrete example of how targeted post-training can shift measured answers while trying to preserve practical ability.
The reusable lesson is evaluation discipline. Start by recording the exact model revision, prompt template, sampling settings and judge. Read the category definitions and examples behind any “bias” score. Test the languages and topics your users actually bring, including ambiguous cases and multi-turn exchanges. Then run your own regression set for the work the model is meant to do: tool calls, retrieval, structured output, accuracy and latency. A model that scores better on a values rubric but breaks your application is not a win.
Nous says it has released the weights, evaluation prompts and training material, which gives others a route to inspect and challenge the result. That openness is valuable. It also leaves the essential question with the deployer: whose values are being encoded, how are disagreements scored, and what happens outside the test set? Arete shows a viewpoint can be changed. Calling that “bias removed” would claim more than the evidence shows.