How to use these docs¶

The SafeTune documentation site is organized to match the library's 2-tier taxonomy. Here is how the UI works and how to find what you need.

Getting Started ──→ Install · quickstart · core concepts (taxonomy)
  │
  ├─ User Guide ───→ Four intervention pillars (harden, recover, unlearn, steer)
  │                  + two instrumentation tools (interpret, evaluate)
  ├─ Reference ────→ API reference · CLI reference · references (papers)
  │                  · feature map · system design · API contract
  ├─ Examples ─────→ Python scripts · notebooks · case study
  └─ Community ────→ FAQ · contributing · scope · changelog

Method group guide pages¶

Each of the six method groups has a guide page like this:

User Guide → Harden
  ├── Overview page      What this pillar does, when to use it
  ├── Gradient surgery   SafeGrad · PlainSFT
  ├── Data shaping       Lisa · SPPFT · LookAhead · STARDSS · DeRTa · CST
  ├── Regularization     AsFT · SAP · Surgery · Booster
  ├── Representation     Vaccine · T-Vaccine
  ├── Tamper-resistant   RepNoise · CTRAP · SEAM · DOOR · MART · ...
  ├── Data selection     SEAL · ConstrainedSFT
  ├── Pre-FT hardening   LoX-Harden
  └── TAR routing        TAR

The left sidebar shows the full tree. Click any method name to see its dedicated page with: - Input/output contract — what you give it, what you get back - Quick example — a runnable code snippet - Parameters — every config field with its default - When to use — which scenarios this method suits - Trade-offs — capability cost, speed, memory

Features¶

Instant search¶

Press s or / or click the search box (top bar). Search is client-side and works offline. Results group by page with highlighted matches.

Dark / light mode¶

Click the sun/moon icon in the top bar. Your preference is remembered across sessions.

Copy code¶

Every code block has a clipboard button (top-right corner). Click to copy the snippet, then paste into your terminal or editor.

Edit this page¶

Every page has an "Edit" link in the top-right action bar. Click it to open the markdown source on GitHub and submit a documentation fix.

Version selector¶

If multiple versions exist, use the version dropdown (top bar) to switch between release docs.

The taxonomy: how methods are organized¶

TIER 1 — INTERVENTIONS  (methods that CHANGE a model's safety)
├── Train-time           harden      — replace your SFT trainer
├── Weight-space         recover     — patch a finished model (no training)
│                        unlearn     — train a model to forget a capability
└── Inference-time       steer       — wrap a frozen model with hooks

TIER 2 — INSTRUMENTATION (methods that OBSERVE safety)
├── Diagnose             interpret   — locate directions / neurons / circuits
└── Measure              evaluate    — run redteam attacks + judge scoring

Each box is a catalog of alternatives. You pick one method per task, not a sequence.

Audit badges¶

Every method in the Feature Map carries a badge:

Badge	Meaning	Can I cite the paper?
Faithful	Implements the cited algorithm exactly	Yes
Simplified	Reduced but algorithmically correct	With caveats
Variant	SafeTune heuristic; not the named algorithm	No (cite as "SafeTune variant")
Wrong	Algorithm is buggy or incorrect	Do not use
Stub	Not yet implemented	—

The Scope & Audit Status page explains the full audit methodology. Only Faithful methods may be cited as the named method from their paper.

Keyboard shortcuts¶

Key	Action
`s` or `/`	Open search
`p` or `,`	Previous page
`n` or `.`	Next page
`Esc`	Close dialog / clear search
`1`–`4`	Switch between tabs (in tabbed content)

Finding what you need¶

If you want to…	Go to…
Install the library	Getting Started → Installation
Understand the taxonomy	Getting Started → Core Concepts
Pick a harden method	User Guide → Harden
See if a method is trustworthy	Reference → Feature Map
Run a quick demo	Examples → Notebooks
See all method papers	Reference → References
Contribute a fix	Community → Contributing
Understand system internals	reference/system-design.md (dev docs)