Machine Learning
Amazon SageMaker Clarify
Amazon SageMaker Clarify analyzes datasets and models for bias and explains supported predictions. For GenAI governance it contributes evidence about data slices and model behavior but does not alone establish fairness or factuality.
Key points
- Pre-training metrics inspect label and feature distributions
- Post-training metrics compare outcomes across configured facets
- SHAP-based explainability applies to supported model and data arrangements
When to use it
- Assess bias across demographic slices before model approval
- Generate feature-attribution evidence for a predictive tool used by an agent
Exam tips
- Choose Clarify for bias and explainability analysis; use task-specific evaluation for generation quality
- Define facets and acceptable differences with domain owners instead of accepting one universal threshold