Notebooks

Interactive Colab notebooks demonstrating every SafeTune workflow.

Demos — one per pillar

One notebook per pillar. Each runs to completion with printed output. Start with steer_demo — no GPU needed, ~2 minutes.

No GPU? Run 01, 02, 05, 09 (steer, recover, interpret, safety monitoring — all inference or light patching, no training). Have a GPU? Everything works, including 10 (full_pipeline), which is the longest single run since it exercises all six pillars back to back.

#	Notebook	Pillar	GPU	Link
01	steer_demo ← start here	Steer	No GPU	OPEN IN COLAB
02	recover_demo	Recover	No GPU	OPEN IN COLAB
03	harden_demo	Harden	GPU helps	OPEN IN COLAB
04	unlearn_demo	Unlearn	GPU helps	OPEN IN COLAB
05	interpret_demo	Interpret	No GPU	OPEN IN COLAB
06	evaluate_demo	Evaluate	GPU helps	OPEN IN COLAB

Comparisons — pick the right method

Side-by-side runs of multiple methods on the same checkpoint.

#	Notebook	Pillar	GPU	Link
07	steer_comparison	Steer	GPU helps	OPEN IN COLAB
08	recover_comparison	Recover	GPU helps	OPEN IN COLAB

Advanced

#	Notebook	Pillar	GPU	Link
09	safety_monitoring	Evaluate	No GPU	OPEN IN COLAB
10	full_pipeline	All pillars	GPU helps	OPEN IN COLAB

Prefer plain scripts?

The six pillar demos (01–06) and full_pipeline (10) each mirror a script in examples/ (examples/quickstart/ or examples/demos/); steer_comparison, recover_comparison, and safety_monitoring (07–09) are notebook-only. Where a script exists, it's the source of truth; the notebook wraps the same code for interactive Colab sessions.