Acceptable use

The short version

Senbonzakura edits open AI models so that they stop refusing. A model edited this way answers requests the original declined, including harmful ones. That is what it does, and everything below follows from it.

Don't put an edited model in front of other people without saying what it is. Not on a website, not in a product, not in a chat bot, not behind an API. The people using it can't see the model's history and have no way to know the refusals are gone.

Don't use the app to produce material that is illegal where you are. Removing a model's refusals doesn't make the output lawful, and it doesn't move responsibility for it onto the model or onto us.

Don't edit a model you aren't allowed to change. See the section on licences below.

Who it is for

Senbonzakura is for people aged 18 and over. A model that has been edited this way will talk about anything, including violence, self-harm and sexual content, with no filter in the way. That isn't something children should be handed, and it isn't something you should hand them either.

Nothing in the app checks your age. This is a condition of use, not a control. If you are under 18, please don't use it. If you are an adult sharing an edited model, the rule above applies twice: say what it is, and keep it away from anyone who shouldn't be reading its answers.

What it is for

• Measuring how much an edit changes what a model can do, which is the part most edits never measure.
• Safety and research work into refusal: where it sits in a model, how fragile it is, and what it takes to remove it.
• Red teaming a model you own or are authorised to test.
• Checking a model someone else published, before you rely on it.

What it is not for

• Running an uncensored service for other people to use without telling them.
• Generating material that is illegal where you are, or that is aimed at a specific real person.
• Removing refusals from a model whose licence forbids it, or whose use policy your edited copy would breach.
• Stripping a safety classifier or a moderation model of its ability to decline.
• Publishing numbers from the app as if they came from somewhere else, or without the measurements behind them.

Licences that follow the models

Two sets of rules usually apply, and they do not merge. The app is ours, under our terms. Each model keeps the licence it came with. Editing a model is a change to it, not a new licence for it: nothing the app does can loosen the terms the original model arrived under.

Several model families carry a use policy that binds anyone who changes the model, and some of those policies speak directly to removing safety behaviour. Read the licence your model came with, and satisfy it before you share the result. The app records the licence of the model you edit. That record is evidence of what the model came with, not proof that you complied with it.

If you share an edited model

• Say that it has been edited, in the description and in the name.
• Carry the original model's licence and any use policy that came with it.
• Keep the measurements behind any number you quote, or say which ones you are keeping back.
• Don't claim a model lost "little" or "no" ability without a measurement to back it up.

Reporting a problem

If you see the app, or a model made with it, used for any of the things above, tell us at [email protected]. If the report is sensitive, say so in the email.

No warranty, and where responsibility sits

The app comes with no warranty of any kind, on its correctness, its fitness for any purpose, or the accuracy of any number it produces. Responsibility for what is done with it, and with any model it edits or measures, sits with whoever does it.

You are the one who runs it, you are the one who decides what to ask the result, and you are the one answerable for what comes out. We are not a party to your use of it, we don't review what you generate, and we don't carry liability for it. To the extent the law where you are allows, that liability is excluded; where it does not, it is limited to the minimum that law permits.