Schismatic

A free software AI model uncensorship tool.

What is Schismatic?

Schismatic is a free software toolkit that takes an existing open-weights language model and surgically removes its built-in β€œI can’t help with that” refusals β€” without retraining the model, changing its facts, or degrading its writing quality.

This Hugging Face organization is the official home of the modified models produced by the Schismatic tool. If you don’t want to run the tool yourself, you can simply download a published model from here and use it directly.

What it does (and doesn’t)

How it works, briefly

  1. Probe β€” feed the model a batch of harmless and harmful prompts and watch which inner signals fire on the harmful ones.
  2. Find the direction β€” identify the single direction inside the model that most strongly predicts a refusal.
  3. Subtract it β€” apply a small, surgical edit that removes that direction from a handful of internal layers.
  4. Measure β€” score the modified model against the original to confirm it still works correctly on normal tasks.

Supported model families

Recipes ship for:

See the benchmarks page for the latest numbers and the studies page for the published hyperparameter searches.

Who it’s for

Researchers, developers, and curious people who want to:

Install the tool

pip install schismatic

Then point it at any supported open-weights checkpoint to produce your own uncensored variant on your own hardware.

Project links

License

Schismatic β€” including all code, recipes, study artifacts, and the modified models published under this organization β€” is licensed under the GNU Affero General Public License, version 3 or later (AGPL-3.0-or-later).