Zentropi — Content Classifiers That Follows Your Policy

We build policy-adaptive content classification models that read your policy and label content against it, instead of forcing you into a fixed taxonomy. Platform teams keep authoring the policies they always have; the classifier follows along.

Our flagship is the CoPE model family (COntent Policy Evaluator): accurate, efficient, single-pass models fine-tuned for policy interpretation. Provide a free-form policy alongside the content, and CoPE returns a label calibrated to that policy.

For a more rigorous understanding of how our models work, check out this research paper: CoPE: A Small Language Model for Steerable and Scalable Content Labeling

What's hosted here

This is the official Zentropi HuggingFace repo. New CoPE model checkpoints and accompanying datasets are released here.

Models

• CoPE-A-9B — first-generation 9B classifier
• CoPE-B-A4B — 2nd-generation text classifier (Apache 2.0)
• CoPE-B-A4B-MM — multimodal variant for text, images, and video (subscriber-only)

For our hosted API and policy authoring platform, visit zentropi.ai.

Learn more

• Website — zentropi.ai
• Skills — github.com/zentropi-ai/skills
• X / Twitter — @zentropi_ai
• Community — ROOST Model Community
• Contact — info@zentropi.ai