API Reference

SafeTune exposes two layers for the same methods:

Layer	Import	Best for
Runner (recommended)	from safetune.runner.<pillar> import <Trainer>	One-line, config-driven training/eval with checkpointing and metric rollup.
Interventions (low-level)	from safetune.<pillar> import ...	Composing a method into your own training loop or research code.

Every method is a self-contained class or function — you pick one per task, you do not chain them. The six pillars:

Pillar	Module	What it does
Harden	safetune.runner.harden	Train-time defenses that keep a model safe while it is fine-tuned.
Recover	safetune.runner.recover	Weight-space repair of a model whose safety has already drifted.
Unlearn	safetune.runner.unlearn	Remove specific harmful capabilities/knowledge.
Steer	safetune.runner.steer	Inference-time activation / decoding steering (no weight changes).
Interpret	safetune.interpret	Locate safety neurons and safety circuits.
Evaluate	safetune.evaluate	Safety benchmarks, red-team attacks, judges, monitoring.

Shared configuration is described on its own page.

Import cheat-sheet

# Runner layer — the common path
from safetune.runner.harden  import SafeGradTrainer
from safetune.runner.recover import ReStaTrainer
from safetune.runner.unlearn import NPOTrainer
from safetune.runner.steer   import CAATrainer

# Interpret / Evaluate are functions + configs
from safetune.interpret import identify_safety_neurons, safety_circuit_info
from safetune.evaluate  import evaluate, AbliterationAttack, TamperBenchEvaluator

Each trainer takes a model (and, where relevant, a tokenizer) plus method-specific keyword arguments, then exposes train(...) / apply(...) / evaluate(...) depending on the pillar. See each pillar page for the exact surface.