---

name: alibi description: "Routes Alibi users to the right explanation, confidence, prototype, or optional-backend workflow and points them to the bundled helpers needed to run it." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

Alibi

Alibi is a Python library for model explanation, confidence scoring, prototype selection, and explanation persistence.

When to use this skill

Use this skill when a user asks how to:

• install Alibi or choose extras
• inspect public Alibi APIs
• run a small reproducible Alibi workflow
• decide which explanation family fits a task
• diagnose MissingDependency, backend, or predictor-shape issues

First reads

• references/workflows.md for the route map
• references/optional-dependencies.md for extras and MissingDependency placeholders
• references/troubleshooting.md for import, backend, and workflow failures
• references/repo-provenance.md when checking staleness or before refresh work
• references/api-reference.md for verified constructor signatures and return shapes

Minimal install and import

• python -m pip install alibi
• From a checkout: python -m pip install -e .
• Minimal import check: python -c "import alibi; print(alibi.__version__)"
• If the import trips over a spaCy / click issue, repair the base environment before assuming the package is broken.

Route map

sub-skills/global-tabular-explanations/

Use for ALE, partial dependence, PD variance, permutation importance, and other global tabular feature-effect workflows.

sub-skills/anchors-local-explanations/

Use for AnchorTabular, AnchorText, and AnchorImage workflows, including text sampling and image segmentation guidance.

sub-skills/attribution-and-shap/

Use for KernelShap, TreeShap, and IntegratedGradients, plus optional SHAP / TensorFlow backend selection.

sub-skills/counterfactuals-and-similarity/

Use for Counterfactual, CEM, CounterfactualProto, CounterfactualRL, and GradientSimilarity workflows.

sub-skills/confidence-prototypes-utilities/

Use for TrustScore, LinearityMeasure, ProtoSelect, and save/load round-trips.

Helper scripts

• scripts/core_smoke.py runs a tiny CPU smoke across the base tabular workflows.
• scripts/check_optional_backends.py reports which optional exports are placeholders and which extra enables them.

What not to do

• Do not treat a MissingDependency placeholder as a verified runtime path.
• Do not send users to the source checkout for docs, examples, or notebooks.
• Do not assume optional backends are installed until the helper says so.