CheckList is a comprehensive behavioral testing framework for NLP models — organizing tests into a matrix of linguistic capabilities × test types (MFT, INV, DIR), providing systematic coverage of model behaviors beyond aggregate accuracy metrics.

CheckList Test Types

• MFT (Minimum Functionality Tests): Simple inputs with known correct answers (e.g., "I love this" → positive sentiment).
• INV (Invariance Tests): Perturbations that should NOT change predictions (e.g., changing a name).
• DIR (Directional Expectation Tests): Perturbations that should change predictions in a known direction.
• Capabilities: Vocabulary, negation, taxonomy, robustness, fairness, temporal, etc.

Why It Matters

• Beyond Accuracy: Accuracy on benchmarks hides capability-specific failures — CheckList exposes them.
• Systematic: The capability × test-type matrix ensures comprehensive, organized model evaluation.
• Actionable: Each failed test pinpoints a specific model weakness to address.

CheckList is the comprehensive test suite for NLP — systematically testing every linguistic capability with targeted, organized test cases.