senbonzakura 0.4.1

Precision abliteration, with receipts.

Documentation  ·  Changelog  ·  What is and is not established

Senbonzakura removes the refusal behaviour from an open-weight language model, and measures what that removal cost.

Removing it is the easy half. Any abliterator can stop a model saying "I can't help with that". The hard part is knowing whether you also took out its reasoning, and a tool that cannot tell the difference will report a lobotomy as a success. So every number here arrives with its conditions: what it was measured on, on rows nothing was fitted or selected on, with an interval, and with a control that would expose it if it were measuring the wrong thing.

Named for Byakuya Kuchiki's zanpakutō, the sword that scatters into a thousand blades.

What it makes. A model that will answer requests the original refused, including harmful ones. That is the point, and it is permanent in the weights. The base model's licence still governs the result. Don't put one in front of other people without saying what it is. The package carries roughly 6,200 harmful prompts so it runs offline.

The detail, and ACCEPTABLE-USE.md ships in the package.

Try it without installing anything

Open the notebook in Colab. Free GPU, nothing on your machine. It measures a model, edits it, then measures what that cost. Nobody has timed it on Colab's hardware, so it gives you no duration to hold us to.

Or, if you have Docker:

docker run --rm ghcr.io/elementmerc/senbonzakura:v0.4.0 doctor

doctor reports what your install can and cannot do, which on a first run is more useful than it sounds.

Install

pip install senbonzakura
senbonzakura setup

That brings torch, transformers, accelerate and optuna with it: 70 packages, 5.9 GB, fifteen of them CUDA wheels. There is nothing to choose.

This paragraph used to tell you not to run that command, and it was right to until 0.4.0 went out. What we got wrong has that one and every other correction, kept rather than deleted.

--track default works from it. The wheel carries the packed evaluation track and six research corpora, so the commands below run straight after installing, with no corpus to fetch and nothing to build. That is also roughly 6,200 harmful prompts written to your disk, which is the thing to know before you install rather than after; the detail is below.

On Linux you get the quantiser too. pip picks the most specific wheel that fits, so a glibc 2.35 or newer machine receives one carrying llama.cpp's binaries and convert and quantise work immediately. Everywhere else, and on older Linux, the universal wheel installs and works without them. senbonzakura doctor says which you got.

senbonzakura setup exists because pip picks by platform, not by hardware. On Windows, PyPI's torch is CPU-only, so a card there sits idle and nothing warns you. It says what it found and prints the command that fixes it, changing nothing unless you add --apply. Install has the full table.

Editing a model wants a CUDA card, and how much of one scales with the model: the weights have to be resident in full precision, because abliteration rewrites them in place. 6 GB is enough for the sizes this project has actually measured, all under 3B. What size card does the arithmetic, including the trap that a mixture-of-experts model needs room for all its experts and not just the active ones. The measuring commands run on CPU.

One command, on your own machine

No corpus, no GPU, no model to edit first. This reads the evaluation track bundled in the install, so it works straight after pip install:

senbonzakura compass \
    --model Qwen/Qwen3-0.6B \
    --harmful default/bad_eval_ds \
    --harmless default/good_ds \
    --skip-harmful 0 --skip-harmless 0 --n 12 \
    --out compass-toy.json --device cpu

Why this model and not a smaller one. This example used HuggingFaceTB/SmolLM2-135M-Instruct until 2026-09-26, which downloads in seconds and then exits 1: at 135M the model does not put a verdict token where the compass reads one, so the run says MARGIN_READOUT_SUSPECT, refuses to call its own AUC a measurement, and is right to. That is the tool being honest, and it is a poor first command, because it looks like a broken install. Qwen3-0.6B answers with a verdict on 100% of prompts and exits 0. It is a larger download, and a first command that works is worth it.

This used to read examples/toy-track/..., which is in the repository and in no install, so the first command under a heading saying "on your own machine" could not be run by anyone who had installed the tool rather than cloned it. Two readers hit it on 2026-09-26 and neither could get past it without guessing. From a clone, examples/toy-track/ still works and is smaller.

This writes two files, and one of them holds harmful prompts in the clear. Beside compass-toy.json you get compass-toy.margins.jsonl, the per-prompt rows the AUC was computed from, and twelve of those rows are the harmful prompts themselves, unwrapped, in whatever directory you ran the command in. They are there so the number can be checked rather than believed, which is the whole argument of this project, and they are the same prompts the wheel deliberately keeps obfuscated. Do not git add that file. Add --no-margins to the command above if you would rather not have it at all.

Said here because this is the first command most people run, and until 2026-09-26 it was said only on two pages this section does not link, so the reader most likely to be surprised was the one least likely to have been told.

Do not quote what it says. Twelve rows measures nothing, and the tool makes you pass those three flags rather than pretending otherwise. The run also prints its own null controls, and on twelve rows one of those will often beat the instrument; if it does, the number is not a weak measurement, it is not one at all.

The real thing

# A corpus with a split that stops you marking your own homework.
senbonzakura track --harmful harmful.txt --harmless harmless.txt --out mytrack

# Search for a configuration and apply it. Timed at 108 minutes end to end on a 6 GB card, at
# the defaults, with the weights already cached, for Qwen3-0.6B: a different model of about the
# same size. The model below has not been timed at the default budget.
senbonzakura kageyoshi --model Qwen/Qwen2.5-0.5B-Instruct --track mytrack --out abliterated --device cuda

# Ask what the edit cost.
senbonzakura compass --model abliterated \
    --harmful mytrack/bad_eval_ds --harmless mytrack/good_ds --out compass.json

harmful.txt and harmless.txt are yours to supply, deliberately. The track page covers building an equivalent from public sources. The split matters more than the prompts.

What it measures

Each is a command, and each reports an interval rather than a bare number:

Two habits run through all of them. The split is three-way, so the rows a configuration is selected on are never the rows it is reported on. And every figure arrives with a control, usually a ruler reading nothing but prompt length: if that separates the arms as well as the real instrument does, the real instrument is measuring sentence length.

The honest ceiling. No measurement above 3B parameters still stands: Qwen3-4B was run in July 2026 and withdrawn, and nothing has been re-run above 3B since. The instruments also weaken as the model does. Qwen3-1.7B's compass scores 0.9887 against a length-only ruler's 0.6564. Qwen3-0.6B scores 0.6616 against that same 0.6564, which is not a measurement of anything.

• What is and is not established — the full account, including the results that went against us.
• REPRODUCING.md — every figure above, mapped to the file it came from and the command that makes it. No GPU or corpus needed to check them.
• METHOD.md — how to measure a behavioural property of a model defensibly. Not about this tool.
• probes/ — add a behaviour of your own for the tool to measure.

Documentation

elementmerc.github.io/senbonzakura

What this repository does not contain

By design, this is methods and results, not a loaded weapon:

• No model weights in this git tree. Abliterated checkpoints are published separately, on the Hub, and they are public and ungated: each one under its base model's own licence, which travels with the weights and is not ours to loosen. The evaluation numbers on most of those cards are withdrawn, for the reasons in the changelog, so read the card before trusting a figure on it. (This bullet twice said the opposite of the truth; see what we got wrong.)

• No harmful prompt sets in this git tree, and this is the bullet that needs the most care. The evaluation track is published separately as a gated dataset under CC BY-NC 4.0. It holds prompts only: no completions, no answers.

  A released wheel is a different matter. It carries roughly 6,200 harmful prompts, wrapped so a scraper does not find them in plaintext. The wrapping is a speed bump, not protection: the key ships beside them and bundled.py says so. To find out whether they are on your disk, run senbonzakura doctor, which lists every bundled corpus it can decode, with its licence.

  This paragraph used to end "neither install you can run today carries them", and explained that 0.3.0 on PyPI predated the bundled track and that a git+... build generates the blobs rather than committing them. That was true when it was written and stopped being true at this release: the wheel carries both blobs, so a plain pip install now puts those prompts on your disk. A sentence telling a reader which installs hold harmful content is a sentence that goes stale silently, and the wrong direction for it to go stale is reassuring. Asking the tool cannot go stale. The dataset card has the channel-by-channel table.

  Where that number comes from, since a figure nobody can derive is a figure nobody can check: 4,895 in the packed evaluation track (259 fitting rows plus 4,636 evaluation rows) and 1,333 across the five harmful research corpora (advbench 520, harmbench 200, harmbench-copyright 100, strongreject 313, xstest-unsafe 200). A sixth corpus, xstest-safe, holds 250 BENIGN prompts used as controls and is not counted here. This said "roughly 6,500" until 2026-09-25, which was only reachable by counting those 250 benign controls as harmful.

• No harmful outputs.

Abliteration removes safety guardrails wholesale. That is both the point and the danger. Use it accordingly.

Disclaimer. This software is provided without 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 modified or measured using it, sits with whoever does it. A model with its refusals removed will answer things a deployed model should not; putting one in front of other people is a decision with consequences that belong to whoever makes it.

Note on licences, and there are three separate ones in play:

• The code is AGPL-3.0-or-later (it embeds a keyword metric copied from Heretic, which is AGPL). If you run a modified version of this code as a network service, the AGPL's section 13 obliges you to offer your modified source to the people using it over that network. That is the clause that distinguishes the AGPL from the GPL, and it is the one that bites an abliteration-as-a-service business rather than an internal user: running it privately, however commercially, triggers nothing. Two readers with no knowledge of this project went looking for this in 2026-09 and found it stated nowhere but the licence file itself, which is the wrong place for the one licence fact a commercial reader most needs.
• The bundled evaluation track under senbonzakura/data/ is a separate work aggregated into the same wheel, and it is CC BY-NC 4.0: non-commercial. The package metadata carries one licence expression and that expression describes the code, so if you are using this commercially, supply your own corpus with --track rather than using --track default. Full attribution is in THIRD-PARTY-NOTICES.md, which is installed beside the package.
• A model you abliterate keeps the base model's licence and use restrictions: redistributing an abliterated checkpoint is governed by that upstream licence (Qwen, Llama, Gemma and so on), not by this repository's.

Credit

• Arditi, Obeso, et al. Refusal in Language Models Is Mediated by a Single Direction (2024). The direction method this builds on.
• Heretic by p-e-w (Philipp Emanuel Weidmann), AGPL-3.0. The automated, KL-guarded search this builds on, and the keyword metric reported here for comparison (its marker list copied verbatim and its normalisation adapted, which is why this project is AGPL; see THIRD-PARTY-NOTICES.md).
• Maxime Labonne. Uncensor any LLM with abliteration. The tutorial that popularised the technique.

Getting help

Ask in Discussions, report a defect in Issues, and report a security problem by email rather than in public: SECURITY.md has the address and what is in scope. SUPPORT.md says what makes a question easy to answer, and what not to paste into a public thread.

Contributing

See CONTRIBUTING.md. Pull requests need a test, a commit message that says why, and one line added to CONTRIBUTORS.md agreeing to the CLA. The project is AGPL and stays AGPL; you keep the copyright in what you write. Everyone taking part is held to the code of conduct.

Licence

Copyright (C) 2026 Daniel Iwugo.

AGPL-3.0-or-later. See LICENSE.

Senbonzakura is a modified work based in part on Heretic, and it is not Heretic. Modified by Daniel Iwugo; first included 2026-07-14, most recently modified 2026-09-23. Only the keyword rate is shared code: its marker list is kept byte-identical and its normalisation is adapted from upstream, so only that number is a like-for-like comparison with Heretic; everything else here is measured by our own instrument. The full statement, and what was and was not changed, is in THIRD-PARTY-NOTICES.md.