Examples

SafeTune ships runnable examples under examples/. They come in three forms:

• Python Scripts — small, self-contained .py files grouped by purpose (quickstart/, demos/, runner/, case_studies/). Most run on CPU with Qwen/Qwen2.5-0.5B-Instruct and are the recommended starting point for integration and CI.
• Notebooks — 10 Colab-ready .ipynb walkthroughs, one per pillar (plus comparison and full-pipeline notebooks) with narration and plots.
• Case study — examples/case_studies/single_direction_jailbreak.py, a fully reproducible, judge-free red-team demonstration that refusal in an aligned model rides on a single direction. Written up in Single-direction jailbreak.

By pillar

SafeTune spans six pillars. Pick the example that matches the mechanism you need:

Pillar	What it does	Quickstart script	Demo / notebook
Harden	Train-time defence (replaces your SFT trainer)	quickstart/harden_quickstart.py	runner/harden.py, harden_demo.ipynb
Recover	Weight-space repair of a drifted model — no training	quickstart/recover_quickstart.py	demos/compare_recover.py, runner/recover.py, recover_demo.ipynb, recover_comparison.ipynb
Unlearn	Remove a capability via forget-set training	quickstart/unlearn_quickstart.py	runner/unlearn.py, unlearn_demo.ipynb
Steer	Inference-time steering — no weight edits	quickstart/quickstart.py	demos/abliteration_and_defense.py, runner/steer.py, steer_demo.ipynb, steer_comparison.ipynb
Interpret	Locate safety circuits and neurons	quickstart/interpret_quickstart.py	demos/interpret_demo.py, demos/run_safety_neuron_inference.py, interpret_demo.ipynb
Evaluate	Measure safety: benchmarks, red-team, drift monitor	quickstart/evaluate_quickstart.py	demos/eval_model.py, demos/evaluate_demo.py, evaluate_demo.ipynb, safety_monitoring.ipynb

The end-to-end demos/full_pipeline.py and full_pipeline.ipynb chain several pillars (Measure → Diagnose → Recover → Verify → Deploy).

Running them

# From the repo root, any script:
python examples/quickstart/quickstart.py

# Override the default model:
python examples/quickstart/quickstart.py --model meta-llama/Llama-3.2-1B-Instruct

See Python Scripts for the full catalogue and per-script run commands, or Notebooks for the Colab-ready walkthroughs.