NousResearch fine-tunes Qwen3.8-27B into Arete-27B to reduce propaganda in responses

The project used a custom dataset of 803 prompts, judged by Grok 4.7, to measure factual accuracy and propaganda alignment. Training data was generated from English-language government sources, with strict deduplication to avoid overlap with the evaluation set. The team emphasizes that Arete-27B is intended for research and users seeking alternative value alignments, not as a definitive fix for all biases. The model’s training pipeline and dataset are open-sourced for further experimentation.