A model can repeat accurate facts yet still adopt a state’s framing of them. This release treats those as separate behaviors: its evaluation scores factual recall and whether answers endorse or substitute an official Chinese-state account. That distinction makes the project useful not just as a model, but as a case study in how supervised fine-tuning can shift a model’s viewpoint.
AI Model2026
Arete-27B
Fine-tunes Qwen3.8-27B to change how it answers politically sensitive questions, targeting both factual recall and framing. The training reinforces values selected by Nous Research; intended for users who want that viewpoint or study value alignment.
Introduction
Information
- Websitehuggingface.co
- OrganizationsNous Research
- Published date2026/11/06
More Items
AI Image2026
Turns text prompts or up to eight reference images into high-resolution images, with support for editing. An eight-step sampler and built-in prompt upsampler streamline generation, though bf16 inference requires substantial GPU memory.
AI Model2026
A sparse MoE base language model for continued training and research, with 280B total parameters and 14B activated per token. Supports up to a 1M-token context for fine-tuning and inference.