Schismatic

About Schismatic

The plain-language version

Big AI models are trained to politely refuse certain kinds of questions. That training is often useful, but it can also stop the model from answering perfectly reasonable questions about sensitive topics like security, medicine, history, or fiction.

Schismatic is an AI model uncensorship tool. It surgically turns down the part of the model that produces those refusals, while leaving everything else alone. The result is an AI model that answers the questions you actually ask it, with the same knowledge and writing quality it had before.

Everything Schismatic produces is open source and freely available. You can install the tool from PyPI, download the published modified models, read the code that produced them, and reproduce the measurements yourself.

The technical version

Schismatic is a directional-ablation toolkit for transformer language models. It implements MPOA-style abliteration with an SSM-repair stage, a multi-objective TPE hyperparameter search using Optuna (optuna.samplers.TPESampler(multivariate=True) with two maximize directions), and a comparison harness that scores against external reference checkpoints under a fixed evaluation configuration.

Outputs are scored on refusal rate (a multilingual keyword scanner and the Qwen3Guard-Gen-4B classifier), KL divergence from the base model over the first k generated tokens, next-token entropy, and a coherence gate that rejects degenerate outputs.

Recipes ship for Qwen, Qwen3, Qwen3-MoE, Qwen3.6, Gemma 3, Gemma 4, GLM-5, and DeepSeek V4 model families. See the benchmarks page for the latest numbers, the studies page for the published hyperparameter searches, and credits for the prior work this builds on.

Project links

License & author

Schismatic is licensed under the GNU Affero General Public License, version 3 or later (AGPL-3.0-or-later).

Author: Jeff Moe <moe@spacecruft.org>
Loveland, Colorado, USA