Senbonzakura

Senbonzakura performs what the project calls ‘multi-direction refusal abliteration’: it maps out the directions in a transformer model’s activation space that encode refusal responses, then orthogonalises those directions out of the model’s weight matrices. The result is a modified copy of the model that the author’s runs show handles previously-refused prompts more consistently than single-direction edits produce. The workflow is local and offline — point the tool at a model checkpoint, run the procedure, receive an edited checkpoint.

The differentiating design choice is the multi-direction cut. The foundational Arditi et al. (2024) paper demonstrated that refusal in transformers is concentrated along a single dominant direction in activation space. Senbonzakura’s position, grounded in the author’s own experimental runs, is that residual refusal survives single-direction ablation because the behaviour is distributed across several directions, not one. By automating direction search via the Heretic approach and then ablating the set of found directions simultaneously, the tool targets what the author describes as the full ‘thousand blades’ of the refusal subspace.

The tool fits squarely in a research and experimentation context: alignment researchers studying refusal mechanisms, developers producing less-censored variants of open-weight models for controlled evaluation, and anyone exploring what refusal looks like geometrically inside transformer activations. It does not fit production pipelines that require auditability, reversibility, or any guarantee that weight edits are scoped only to refusal behaviour — the orthogonalisation procedure touches weight matrices, and the docs do not describe a formal analysis of collateral effects on unrelated model capabilities.

Senbonzakura is released under the AGPL-3.0 license, is self-hosted only, and carries no API surface or commercial tier. It is a Python package (see pyproject.toml) targeting mid-sized open-weight LLMs; the repository is early-stage with nine commits at the time of curation and a fork/star count that reflects pre-community-adoption status.