Nous Research releases Arete-27B, a post-trained Qwen3.8-27B with reduced state-aligned bias

• Nous Research released Arete-27B, an open-weight model post-trained on Alibaba's Qwen3.8-27B to remove state-aligned biases and reinforce Western values.
• On a custom 803-prompt evaluation, the model cut the rate of answers containing propaganda from 22% to 1% and raised the fact score from 2.32 to 2.98 out of 5.
• General capabilities remained stable, with the model scoring 65.60% on Hermes Bench and 70% on Terminal-Bench-2.1, showing no significant regression from the base model.
• The project demonstrates that value alignment in open models is a shallow layer that can be modified with approximately 52.8 million tokens of fine-tuning without affecting underlying intelligence.

Methodology and Data

• Training used 20,000 generated prompts across seven value categories, with teacher responses provided by the Inkling model.
• The fine-tuning process involved 1,202 steps over two epochs using AdamW optimizer and bf16 precision, with no reinforcement learning or architectural changes.
• An LLM judge (Grok 4.7) evaluated factual recall and framing, requiring quoted evidence for any propaganda score.

Availability

• Model weights, the Propaganda_Eval dataset, and training data pairs are available on Hugging Face.
• The evaluation scripts are open-source, allowing the community to test other models for similar bias patterns.