Explaining model outputs: MLA-C02 study notes and practice questions
What to know about Explaining model outputs for the MLA-C02 exam, from the ML Model and Foundation Model (FM) Development domain (24% of the exam guide), then 1 free question to test it. Part of the MLA-C02 study guide.
These questions ask how to explain a model’s decisions to regulators, customers or the team, with the least custom work.
SageMaker Clarify explanations
SHAP (Shapley values) assigns each feature a contribution to a prediction.
Local explanations cover one prediction (“why was this loan declined?”).
Global explanations aggregate across many predictions (“which features matter most overall?”).
Partial dependence plots (PDP) show how predictions change as one feature changes.
Reports are written to S3 and shown in SageMaker Studio.
Online explainability returns SHAP values with each real-time endpoint response.
Explanations for images (which regions mattered) and text (which tokens) are supported.
Clarify also evaluates foundation models for accuracy, toxicity and stereotyping.
Monitoring explanations
Model Monitor feature attribution drift compares production SHAP values with a baseline and alerts when the model starts relying on different features.
Choosing interpretable models
Linear models and small decision trees are explainable by design.
Complex models (gradient boosting, deep learning) usually need SHAP or similar post-hoc methods.
Regulated decisions may require both a clear model and documented explanations.
Documenting
SageMaker Model Cards record intended use, training data, evaluation, risks and explanations.
AWS AI Service Cards describe AWS’s own AI services.
Generative AI
Citations in RAG answers show which sources supported the response.
Agent traces show each reasoning step and tool call.
Contextual grounding checks score whether an answer is supported by the sources.
Common traps
Pre-training bias metrics to explain predictions. Use SHAP on the trained model.
Removing features to make a model “explainable” when attributions answer the question.
Writing explanations by hand for thousands of decisions.
A regional bank uses a gradient-boosted model to decide small-business loan applications, and it already runs SageMaker Clarify processing jobs for its bias reports. Examiners require the bank to explain, for each declined application, which features most influenced that decision, using one method across the portfolio.
The explanations must be produced from the trained model and the bank's historical applications in S3, and they must be reviewable as reports, not ad hoc notebook output.
Which approach meets these requirements?
Choose one answer.
AA. Use a Clarify processing job with SHAP explainability on the trained model and the application data, and store the attribution report in S3.
BB. Read the model's feature importance values from a notebook after training, and copy the top features into the examiner memo by hand.
CC. Run a Clarify job that computes the pre-training bias metrics on the application data, and use those metrics as the explanation for each declined applicant.
DD. Turn off the model's feature inputs and use only the loan amount, so the decisions become simple enough to explain without any tool.
Show the answer and explanation
Answer: A
A. Right: Correct. Clarify computes SHAP-based feature attributions for the model and dataset in a managed processing job, and writes a report to S3 that reviewers can use. This gives per-decision explanations at portfolio scale. Learn more: Model explainability with SageMaker Clarify (opens in a new tab)
B. Wrong: Manual copying does not produce a repeatable, portfolio-wide report, and the stem asks for reports rather than notebook output. The method also would not show per-decision attributions consistently.
C. Wrong: Pre-training bias metrics describe the dataset, not the attributions for each decision. They do not explain which features influenced a particular decline.
D. Wrong: Removing features changes the model and may reduce its accuracy, and it does not produce explanations for the existing decisions. The examiners asked for explanations of the current approach.
Explanation
Clarify computes SHAP-based feature attributions from the trained model and data, and writes a report to S3, which gives reviewable explanations across the portfolio. Manual notes, pre-training metrics and removing features each fail to explain the existing decisions in a repeatable way.
Want more on Explaining model outputs? The 5 full-length practice sets have 4 more Explaining model outputs questions, 325 in all, none of them in the free exam. One payment of $14.99, no subscription. See the practice sets →