User Guide

Usage guides for all six method groups — four intervention pillars and two instrumentation tools. Each guide covers the input contract, a quick example, and the catalog of methods.

◆ SafeTune audited LLM-safety methods

Tier 1 · Interventions ↓

Tier 2 · Instrumentation ↓

TIER 1 · INTERVENTIONS CHANGE model safety

hardenTrain-time · base + FT data

recoverWeight-space · drifted model

unlearnForget-set · remove capability

steerInference-time · frozen model

TIER 2 · INSTRUMENTATION OBSERVE safety

interpretDiagnose · locate safety

evaluateMeasure · red-team + eval

• Harden

  ---

  Keep safety during fine-tuning. A Harden trainer replaces your SFT trainer — it IS the fine-tuning.

• Recover

  ---

  Restore safety in a drifted model. Weight-space editing, no training. Methods across 6 granularities.

• Unlearn

  ---

  Remove a capability from a finished model. Forget set + retain set.

• Steer

  ---

  Refuse harmful prompts at inference time. Wraps a frozen model, no weight changes.

• Interpret

  ---

  Locate where safety lives — directions, neurons, circuits. Feeds Steer and localization-aware Recover.

• Evaluate

  ---

  Measure safety with red-team attacks and benchmark/judge eval.

Tips & best practices

1. Pick one method per task. SafeTune is a library of alternatives, not a pipeline. recover, harden, and steer solve different problems by different mechanisms — choose the one that matches your situation.
2. Safety dataset: BeaverTails works well for English LLMs; customize for other languages.
3. Benchmarks: HarmBench + XSTest cover most cases; add custom benchmarks as needed.
4. Steering strength: The strength parameter (named strength in CAA and refusal-direction steering, default 1.0, or multiplier in SCANS/STA/AdaSteer, default 3.0-3.5) controls how aggressively the model is steered. Increase for stronger safety enforcement; watch for fluency cost at very high values.
5. Memory: SafeGrad requires two forward passes (task + safety); use gradient accumulation to fit.