Learning Interpretability Tool

When it fits

Use LIT when you need an interactive interface for debugging or comparing model behavior, rather than only inspecting predictions in code or in aggregate reports.

Who it is for

• Machine-learning practitioners investigating model errors and individual predictions.
• Researchers comparing model outputs or testing how input changes affect behavior.
• Developers integrating custom models and datasets through LIT's APIs.

Documented uses

• Identify types of examples on which a model performs poorly.
• Inspect an individual prediction using salience maps and prediction visualizations.
• Compare models or compare a model's outputs across examples.
• Explore model behavior across data slices, bins, custom metrics and embedding spaces.
• Create edited or generated counterfactual examples and evaluate model responses.