AIPolitics

NousResearch/Arete-27B · Hugging Face (Qwen finetune)

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 07/11/2026, 04:47:25 UTC.

Briefings

NousResearch released Arete-27B, a model post-trained on Alibaba's Qwen3.8-27B to align with western values. Evaluation showed a 2.66-point increase in fact score (2.32 to 2.98) and a 99.1% reduction in propaganda answers (22% to 0.9%). The model maintained agentic capabilities across seven value categories, though the study explicitly notes it did not remediate all behaviors associated with the original model's viewpoint.

Key points

• Arete-27B is a Qwen3.8-27B model post-trained by NousResearch to reinforce western values on a Chinese-trained base.
• The model improved its fact score from 2.32 to 2.98 and reduced propaganda answers from 22% to 0.9%.

Context

NousResearch developed Arete-27B by post-training Qwen3.8-27B to address observed biases in politically sensitive topics. The project aims to provide labs with tools to mold intelligence to varied preferences while ensuring users can still choose from multiple value options.

Limitations

Evaluation was limited to seven value categories and did not cover all behaviors associated with the original model's viewpoint. The study explicitly states it did not remediate all behaviors, and the evaluation checkpoint represents a specific step in the process.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 07/11/2026, 04:48:22 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeigdt5oqjej5v4tg6sti5t5zm5pjcyhlfqc5i2nyj24mm7j55trami