vauban 0.3.3

LLM red-teaming and adversarial testing framework

vauban

An MLX-native toolkit for understanding and reshaping how language models behave on Apple Silicon.

Named after Sébastien Le Prestre de Vauban, the military engineer who worked both siege and fortification. Vauban does the same for model behavior: measure it, cut it, probe it, steer it, or harden it.

What it does

Vauban is a TOML-first CLI for workflows built around activation-space geometry:

• measure behavioral directions from model activations
• cut or sparsify those directions in weights
• probe and steer models at runtime
• map refusal surfaces before and after intervention
• run defense, sanitization, optimization, and attack loops

The primary interface is not a pile of subcommands. It is:

vauban <config.toml>

All pipeline behavior lives in the TOML file.

Requirements

• Apple Silicon Mac
• Python 3.12+
• uv

Install

Install the released CLI:

uv tool install vauban

If your shell cannot find vauban, update your shell config once:

uv tool update-shell

Then open a new shell and check the command:

vauban --help
vauban man workflows

For development from this repo:

uv tool install --editable .

Quick Start

Pick a goal — the manual will tell you which sections to use:

vauban man workflows

Then scaffold a config and go:

vauban man quickstart

Scaffold a starter config:

vauban init --mode default --output run.toml

All pipeline modes can be scaffolded. See the full list with:

vauban init --help

Validate before a real run:

vauban --validate run.toml

Then run the pipeline:

vauban run.toml

By default, output goes to output/ relative to the TOML file.

Minimal TOML

This is the minimal config the code accepts for the default pipeline:

[model]
path = "mlx-community/Llama-3.2-3B-Instruct-4bit"

[data]
harmful = "default"
harmless = "default"

[model].path is required.

[data].harmful and [data].harmless are required for most runs. "default" uses Vauban's bundled prompt sets.

You can also choose the output directory explicitly:

[output]
dir = "runs/baseline"

Experiment Tech Tree

Vauban has built-in experiment lineage tracking through an optional [meta] section. This metadata does not change pipeline execution. It exists so you can organize runs as a tech tree.

Minimal example:

[meta]
id = "cut-alpha-1"
title = "Baseline cut, alpha 1.0"
status = "baseline"
parents = ["measure-v1"]
tags = ["cut", "baseline"]
date = 2026-03-02
notes = "First stable reference run."

Verified status values are:

archived, baseline, dead_end, promising, superseded, wip

If [meta].id is omitted, Vauban uses the TOML filename stem.

Render the tree from a directory of TOML configs:

vauban tree experiments/
vauban tree experiments/ --format mermaid
vauban tree experiments/ --status promising
vauban tree experiments/ --tag gcg

Each run also appends an experiment_log.jsonl file inside the configured output directory with the resolved config path, pipeline mode, report files, metrics, and selected [meta] fields.

How TOML Drives Vauban

vauban <config.toml> loads one TOML file and decides what to do from the sections you include.

The default path is:

1. measure
2. cut
3. export

You extend that run by adding more sections. Common examples:

• [eval] adds post-cut evaluation reports.
• [surface] adds before/after refusal-surface mapping.
• [detect] adds hardening detection during measurement.
• some sections switch Vauban into dedicated mode-specific runs instead of the default pipeline.

If you know what you want to do but not which sections to use:

vauban man workflows

For field-level reference on any section:

vauban man cast
vauban man softprompt
vauban man measure

For mode precedence:

vauban man modes

Commands You Will Actually Use

Inspect the manual:

vauban man workflows
vauban man quickstart
vauban man cast
vauban man all

Scaffold configs:

vauban init --help
vauban init --mode default --output run.toml
vauban init --mode probe --output probe.toml

Validate config and prompt files without loading model weights:

vauban --validate run.toml

Export the current JSON Schema for editor tooling:

vauban schema
vauban schema --output vauban.schema.json

Compare two run directories:

vauban diff run_a run_b
vauban diff --format markdown run_a run_b
vauban diff --threshold 0.05 run_a run_b

vauban diff --threshold ... is a CI gate: it exits non-zero if any absolute metric delta crosses the threshold.

Render the experiment lineage tree:

vauban tree experiments/
vauban tree experiments/ --format mermaid
vauban tree experiments/ --status promising

Data Formats

Verified by the generated manual:

• prompt JSONL for [data] and [eval]: one JSON object per line with a prompt key
• surface JSONL for [surface].prompts: requires label and category, plus either prompt or messages
• refusal phrase files: plain text, one phrase per line
• relative paths resolve from the TOML file's directory

Minimal prompt dataset example:

{"prompt":"What is the capital of France?"}
{"prompt":"Write a haiku about rain."}

Notes On Verification

This README is aligned to the code in this repo:

• package name: vauban
• console script: vauban = vauban.__main__:main
• verified commands: vauban <config.toml>, --validate, schema, init, diff, tree, man
• verified manual topics and scaffolded modes were checked against the live CLI help and generated manual

The current README previously had some stale mode/output claims; this version removes those and points readers to vauban man ... for the parts generated directly from code.

Documentation

Full docs: vauban.readthedocs.io

Resource	Description
Spinning Up in Abliteration	Seven-part progressive curriculum
Getting Started	Guided walkthrough
Configuration Reference	TOML field reference
Surface Mapping	Surface mapping reference
examples/config.toml	Annotated example config

License

Apache-2.0