Changelog

v1.0.0 — Initial public release

SafeTune ships a broad roster of LLM-safety methods, organized into a 2-tier taxonomy and faithfulness-audited against their cited papers.

Methods included

Pillar	Methods	What it does
Harden	27	Training-time defenses — keep safety intact through fine-tuning
Recover	26	Training-free weight-space recovery for drifted models
Unlearn	6	Capability removal via forget-set training
Steer	19	Inference-time intervention without weight changes
Interpret	6	Refusal-direction extraction, safety-neuron localization, circuit discovery
Evaluate	24	Red-team attacks and benchmark scoring

Configuration and extensibility

• SafeTuneConfig — declarative YAML config for CLI runs (SafeTuneConfig.from_yaml), plus the --config CLI flag.
• --train-dataset / --train-split CLI flags — training data fully configurable.
• safetune.runner._registry — centralised algo registry; register_harden() / register_recover() / register_unlearn() let third-party code extend the method menu at runtime.

Audit summary

Every method was audited against its cited paper and carries a badge: most are Faithful, a few are Simplified-correct or Variant (they run correctly but must not be cited as the named method). See the Feature Map for per-method verdicts and Scope & Limitations for what each verdict means.

Getting started

pip install safetune

See Getting started for a 60-second quickstart, User Guide for per-pillar usage, and Notebooks for Colab notebooks.