Zentropi CoPE

weights

open(zentropi-ai/cope-a-9b on HF)

base

Gemma-2-9B(LoRA)

license

zentropi-openrail-m(OpenRAIL-M, use-restricted, NOT OSI)

Zentropi publishes CoPE's LoRA adapter for self-hosting on the Gemma 2 base, behind a click-through form that asks for contact information. Its OpenRAIL-M license prohibits uses such as surveillance and is not OSI-approved, which makes this an open-weights release rather than open source.

• https://huggingface.co/api/models/zentropi-ai/cope-a-9b recorded 2026-09-29

  The Hugging Face API record for zentropi-ai/cope-a-9b lists the file "adapter_model.safetensors" among its siblings, carries the tag "base_model:finetune:google/gemma-2-9b", and records the license as "zentropi-openrail-m" under a custom license tag "license:other", with gated set to auto (a click-through contact-information gate rather than a private repo).

• https://huggingface.co/zentropi-ai/cope-a-9b recorded 2026-09-29

  The model page header says "You need to agree to share your contact information to access this model", the card tells developers to "Merge with the cope adapter (i.e., this model)" on the Gemma base, lists "Surveillance applications" under Prohibited Uses alongside "Other restrictions as enumerated in our", which links to the openrail-m license, and says "As an open model, CoPE can also be self-hosted on your infrastructure".

Zentropi CoPE is a ROOST Model Community partner model, distributed through the RMC. It ships as a LoRA adapter merged into a Gemma base rather than pulled directly, so no meaningful download count exists, and Zentropi publishes no other usage figure.

• https://huggingface.co/zentropi-ai/cope-a-9b recorded 2026-10-08

  Card: "Downloads last month - Downloads are not tracked for this model"; one Space (roosttools/llm_moderation_testing) uses it; no usage figure is published.

Like gpt-oss-safeguard, CoPE evaluates content against an arbitrary policy the user writes rather than a fixed taxonomy, and its card reports F1 results for hate speech, toxicity, sexual content, self-harm and harassment. It returns only a binary label, which puts it level with gpt-oss-safeguard rather than ahead.

• https://arxiv.org/abs/2512.18027 recorded 2026-09-29

  The arXiv abstract for CoPE describes "a policy-steerable small language model capable of fast and accurate content labeling" and says "When evaluated across seven different harm areas, CoPE exhibits equal or superior accuracy to frontier models at only 1% of their size".

• https://huggingface.co/zentropi-ai/cope-a-9b recorded 2026-09-29

  The model card lists "Policy-adaptive content evaluation" and "Steerable (no fixed taxonomy/definitions)", says the model "can accept any arbitrary content policy", describes output as "Binary classification only (i.e., presence/absence of matching labels)", and gives F1 tables headed "Hate Speech Classification", "Toxic Speech Classification", "Sexual Content Classification", "Self-Harm Content Classification" and "Harassment Classification". Searched the card for violence and violent: none.