Free AWS Certified Machine Learning Engineer - Associate Practice Questions

Sample Practice Questions

Try these 10 sample questions from the AWS Certified Machine Learning Engineer - Associate practice bank. Each includes the answer and explanation. Continue with the interactive practice above to review the full question set.

1. 1A fraud model needs low millisecond feature reads for live predictions and historical feature values for retraining. Which SageMaker Feature Store configuration meets both requirements?

   1. A. An online store alone for both workloads
   2. B. An offline store alone for both workloads
   3. C. A feature group with online and offline stores
   4. D. A feature group with neither store enabled

   Answer: C

   Explanation: AWS states, "The online store retains only the latest records for your features." It also states, "The offline store keeps all records for your features as a historical database." Both stores are needed to combine low latency serving with historical training data.

   Reference: Create, store, and share features with Feature Store — Create, store, and share features with Feature Store > store options

2. 2A team must answer questions using proprietary documents retrieved at query time. It wants the retrieved passages to supply context for the generated response. Which modeling approach fits this requirement?

   1. A. Supervised fine-tuning on response style examples
   2. B. Distillation from a teacher into a student model
   3. C. Few-shot prompting with task demonstrations alone
   4. D. Retrieval Augmented Generation using a knowledge base

   Answer: D

   Explanation: AWS states, "When a query is made, a knowledge base searches your data to find relevant information to answer the query." RAG supplies relevant proprietary passages at query time. Style fine-tuning, distillation, and few-shot demonstrations alone do not perform the required document retrieval.

   Reference: Retrieve data and generate AI responses with Amazon Bedrock Knowledge Bases — Retrieve data and generate AI responses with Amazon Bedrock Knowledge Bases > overview

3. 3A steady-traffic application needs interactive synchronous predictions from a warmed, provisioned SageMaker endpoint. It cannot tolerate inference cold-start delays. Which inference option directly fits this operating pattern?

   1. A. Batch transform
   2. B. Asynchronous Inference
   3. C. Real-time inference
   4. D. On-demand Serverless Inference

   Answer: C

   Explanation: AWS states, "Real-time inference is ideal for inference workloads where you have interactive, low latency requirements." The already provisioned, warmed endpoint meets the synchronous operating pattern. Batch and asynchronous modes do not provide the requested interactive synchronous service, and on-demand serverless is intended for workloads that can tolerate cold starts.

   Reference: Deploy models for inference — Additional options > Real-time inference

4. 4An existing Model Monitor customer wants to detect changes in tabular inference input distributions relative to training data. Which monitoring type should it use?

   1. A. Model quality monitoring of prediction accuracy
   2. B. Feature attribution drift monitoring
   3. C. Bias drift monitoring of predictions
   4. D. Data quality monitoring

   Answer: D

   Explanation: AWS lists "Data quality - Monitor drift in data quality." It describes comparing current data with training-data baseline constraints. The other monitoring types concern accuracy, attribution, or prediction bias rather than input distribution quality.

   Reference: Data and model quality monitoring with Amazon SageMaker Model Monitor — Data and model quality monitoring > monitoring types and baseline workflow

5. 5An ingestion application must push updated feature values to SageMaker Feature Store as soon as each update is detected. Which API operation should it call?

   1. A. PutRecord
   2. B. GetRecord
   3. C. DeleteRecord
   4. D. DescribeFeatureGroup

   Answer: A

   Explanation: AWS states, "You can push records to Feature Store by calling the synchronous PutRecord API call." PutRecord writes the updates; reading, deleting, or describing resources does not ingest the new feature values.

   Reference: Create, store, and share features with Feature Store — Feature data ingestion

6. 6A team has labeled input-output examples for a specific task and wants to change a supported foundation model's parameters to learn the desired responses. Which customization strategy matches this goal?

   1. A. Retrieval Augmented Generation
   2. B. Supervised fine-tuning
   3. C. Prompt-only few-shot examples
   4. D. Vector similarity indexing

   Answer: B

   Explanation: AWS describes supervised fine-tuning as "Provide labeled data to train a model to improve performance on specific tasks." It also states that model parameters are adjusted. Retrieval, prompt-only examples, and vector indexing do not themselves train those parameters.

   Reference: Customize your model to improve its performance for your use case — Supervised fine-tuning

7. 7A supported small model receives sporadic requests separated by idle periods. Cold starts are acceptable, and the team wants SageMaker to manage the underlying inference infrastructure. Which endpoint option is designed for this pattern?

   1. A. Serverless Inference
   2. B. A continuously provisioned real-time endpoint
   3. C. A continuously provisioned multi-model endpoint
   4. D. An inference pipeline on fixed instances

   Answer: A

   Explanation: AWS states that Serverless Inference "is ideal for workloads which have idle periods between traffic spurts and can tolerate cold starts." It directly matches the stated traffic and management preferences; continuously provisioned alternatives do not supply that serverless operating model.

   Reference: Deploy models for inference — Additional options > Deploy models with Amazon SageMaker Serverless Inference

8. 8An existing Model Monitor customer captures classification predictions but wants to calculate production accuracy after actual outcomes become available. What additional data must the model quality monitoring job merge with those predictions?

   1. A. Actual outcome labels for the captured predictions, stored in S3
   2. B. The captured predicted classes reused as their own outcome labels
   3. C. The captured confidence scores converted into pseudo-labels
   4. D. Labels from training records without outcomes for the captured predictions

   Answer: A

   Explanation: AWS states that model quality monitoring "merges data that is captured from real-time or batch inference with actual labels that you store in an Amazon S3 bucket". Current actual outcomes determine production correctness. Predictions or confidence-derived pseudo-labels are not ground truth, and training-only labels do not establish the outcomes of these captured production records.

   Reference: Model quality monitoring — Model quality monitoring > overview

9. 9Which preparation satisfies the schema prerequisite before the first ingestion into a new SageMaker Feature Store feature group?

   1. A. Define feature names and let the first ingested record supply their data types
   2. B. Define data types and let field positions supply feature names during ingestion
   3. C. Send the first record and let PutRecord create its feature group and definitions
   4. D. Define the feature group and specify the name and data type of each feature

   Answer: D

   Explanation: AWS states, "To ingest features into Feature Store, you must first define the feature group and the feature definitions (feature name and data type) for all features that belong to the feature group." Both names and types must be defined before ingestion. Deferring either to incoming records or assuming PutRecord creates the group omits this prerequisite.

   Reference: Create, store, and share features with Feature Store — Create feature groups

10. 10A team wants a smaller student foundation model to learn a larger teacher model's behavior on use-case-specific prompts. Which approach implements Bedrock model distillation?

    1. A. Generate teacher responses and use them to fine-tune the student
    2. B. Generate student responses and use them to fine-tune the teacher
    3. C. Add teacher responses only to prompts at student inference time
    4. D. Fine-tune the student only on responses it generated for itself

    Answer: A

    Explanation: AWS states, "Amazon Bedrock generates responses from the teacher model for the given prompts, and then uses the responses to fine-tune the student model." The transfer runs from teacher responses into student training. Reversing it trains the teacher, runtime-only examples do not train the student, and student-only responses do not transfer teacher behavior.

    Reference: Customize your model to improve its performance for your use case — Distillation