What it does

explainerdashboard wraps your fitted model and spits out a Plotly Dash web app showing SHAP values, permutation importances, partial dependence plots, and performance metrics. It handles the data wrangling and rendering so you don’t have to. You can run it in a notebook, export static HTML, or deploy it via YAML config—including auto-detected SageMaker Studio proxy prefixes.

The interesting bit

The library is explicitly designed for the awkward handoff between data scientists and everyone else. It lets you group one-hot encoded features, attach human-readable descriptions, and even build an ExplainerHub that hosts multiple model dashboards in one place. The “simple” mode trims the UI to a single page for less technical audiences.

Key highlights

• One-liner launch: ExplainerDashboard(ClassifierExplainer(model, X_test, y_test)).run()
• Supports sklearn, XGBoost, CatBoost, LightGBM, and skorch (tabular PyTorch)
• Pipeline-aware: auto-strips prefixes like num__Age → Age, infers categorical groups
• Export to static HTML or deploy via YAML for CI/CD
• ExplainerHub for multi-model hosting with titles and descriptions

Caveats

• Requires y_test for performance tabs; without it, those sections simply don’t render
• Categorical grouping for pipelines requires binary-like encoded columns, not arbitrary post-processing
• The README notes the library is modular for custom layouts, but the actual custom-building docs are offsite at readthedocs

Verdict

Grab this if you need to put model explanations in front of managers, regulators, or customers without building a frontend team. Skip it if you’re already deep in a bespoke Streamlit or Gradio setup that covers the same ground.