Responsible use

What this index is

A catalogue of models that are already public. Every entry links to a Hugging Face repository that anyone can find without this site; nothing here is private, leaked, or otherwise unavailable. What this site adds is organisation: which distinct models exist, what each was built from, how it was modified, and what that implies for someone evaluating it.

The audience is professional. Abliterated models are a standard tool for red teams and safety researchers, because you cannot characterise what an unaligned model will produce by testing a model that refuses. Concretely, they are used to build red-team corpora, to develop and validate detection for AI-generated attack content, to establish a defensible baseline for what an adversary holding open weights can actually achieve, and to separate a model's alignment layer from its underlying capability.

Handling controls

These weights have no refusal behaviour and no content filter. Treat them as untrusted, unmoderated code that also emits unmoderated text:

• Run them isolated — no unsandboxed tool or shell access, no production credentials, no outbound network from the inference host.
• If anything beyond your own team can reach the model, put moderation in front of it. The model will not decline on your behalf, so an unmoderated public endpoint is your liability, not the publisher's.
• Verify what you download. These are third-party artifacts from largely unvetted accounts: check hashes, prefer safetensors over pickle-based formats, and scan before loading.
• Log and retain your evaluation prompts and outputs. If you are doing this as authorised work, the record is what demonstrates that.

Obligations ablation does not remove

Removing a refusal direction changes the model's behaviour, not your legal position. The upstream licence — Llama Community Licence, Gemma Terms, Apache 2.0, whatever the base model carries — still governs the weights, including its acceptable-use clauses. A re-release that declares a more permissive licence than its parent does not make that licence apply. Applicable law, your own organisation's policy, and the terms of any engagement you are operating under all continue to bind you.

If you are testing systems you do not own, you need authorisation in writing first. That is true regardless of which model you use to do it.

Corrections and removal

Metadata here is derived from public API data and publisher model cards and may be wrong; see the known limitations. If you publish a model listed here and its entry is inaccurate, or you would like it delisted, that request will be honoured — this index has no interest in misrepresenting anyone's work.