MLA-C02 Practice Test

The AWS Certified Machine Learning Engineer – Associate (MLA-C02) is a associate-level AWS certification. The exam has 85 questions, runs 170 minutes, costs $75 (beta), and requires a passing score of 720/1000. CertSharp provides 500 exam-style MLA-C02 practice questions — 30 free with no credit card — each with a detailed explanation.

1. Question 2Data Preparation for ML and AIEasy

   A data analyst wants to profile a dataset in S3 and clean it visually, without writing code, using prebuilt transformations and data quality rules. Which AWS service is designed for this?

   • A

     Amazon Kinesis Data Streams

   • B

     AWS Glue DataBrew

   • C

     Amazon SageMaker Ground Truth

   • D

     Amazon Macie

   Show answer & explanation

   Correct answer: B — AWS Glue DataBrew

   AWS Glue DataBrew is a visual data preparation tool with profile jobs, prebuilt transformations and data quality rules. Kinesis Data Streams ingests streaming records, Ground Truth coordinates data labeling, and Macie discovers sensitive data in S3.

2. Question 3Operating, Monitoring, and Securing ML and AI SolutionsEasy

   A RAG application sometimes returns answers that are not supported by the retrieved documents. Which Bedrock Guardrails capability detects these hallucinated responses?

   • A

     The hate speech content filter

   • B

     The denied topics policy

   • C

     The contextual grounding check

   • D

     The word filter policy

   Show answer & explanation

   Correct answer: C — The contextual grounding check

   The contextual grounding check compares a response with the reference source and the query, and flags content that is not grounded or relevant. Content filters target harmful categories such as hate. Denied topics block subject areas, and word filters match specific words, so none of them judge factual support.

3. Question 4Deployment and Orchestration of ML and AI WorkflowsEasy

   A developer builds a multi-turn chat on a Bedrock agent. Each follow-up question is answered as if the conversation just started. What should the developer change when calling InvokeAgent?

   • A

     Pass a new randomly generated sessionId on each call so the agent starts a clean context for each turn

   • B

     Pass the same agentAliasId on each call and leave the sessionId as a new value on each request

   • C

     Pass the same sessionId on each call so the agent keeps the conversation context across turns

   • D

     Pass the enableTrace flag on each call so that the agent stores earlier turns in its context window

   Show answer & explanation

   Correct answer: C — Pass the same sessionId on each call so the agent keeps the conversation context across turns

   Conversation history is tied to the sessionId, so reusing it lets the agent see earlier turns. A new session ID resets context, the alias identifies the agent version rather than the conversation, and enabling trace just returns reasoning details for debugging.

4. Question 5ML Model and Foundation Model (FM) DevelopmentEasy

   A team wants a record of the prompts and responses sent to Amazon Bedrock models for monitoring of a RAG application. Which feature provides this?

   • A

     SageMaker Clarify explainability reports on the Bedrock model

   • B

     SageMaker Ground Truth labeling jobs for each response

   • C

     SageMaker Model Monitor data capture on the Bedrock model

   • D

     Bedrock model invocation logging delivered to Amazon CloudWatch Logs or Amazon S3

   Show answer & explanation

   Correct answer: D — Bedrock model invocation logging delivered to Amazon CloudWatch Logs or Amazon S3

   Bedrock model invocation logging can deliver request and response data, along with metadata, to CloudWatch Logs or S3 for monitoring and analysis. Clarify explains SageMaker AI models. Ground Truth labels data with humans. Model Monitor data capture applies to SageMaker AI endpoints, not Bedrock model invocations.

5. Question 6Data Preparation for ML and AIMedium

   A team trains models with SageMaker Pipelines and wants each run to stop automatically when the training data shows bias beyond agreed thresholds for a sensitive facet. Which pipeline addition meets this requirement?

   • A

     A ClarifyCheckStep configured for data bias that compares metrics against a baseline and fails the run on violations

   • B

     A QualityCheckStep for data quality, which compares feature statistics but does not compute facet bias metrics

   • C

     A ConditionStep that checks the training job's validation accuracy before the model registration step runs

   • D

     A TransformStep that scores the training data and writes predictions to S3 for later manual bias review

   Show answer & explanation

   Correct answer: A — A ClarifyCheckStep configured for data bias that compares metrics against a baseline and fails the run on violations

   ClarifyCheckStep runs a Clarify check inside the pipeline, including data bias checks, and can be set to fail the execution when a violation of the baseline constraints is found. QualityCheckStep profiles statistics and drift and does not compute facet bias metrics. A ConditionStep on validation accuracy does not look at facet-level bias. A TransformStep with manual review adds no automatic enforcement.

6. Question 7Operating, Monitoring, and Securing ML and AI SolutionsMedium

   A Bedrock agent fails to invoke the Lambda function behind its action group and reports a permissions error. The Lambda execution role is correctly configured. What is the fix?

   • A

     Add lambda:InvokeFunction to the Lambda execution role so the function can call itself

   • B

     Add a resource-based policy on the Lambda function that lets bedrock.amazonaws.com invoke it for that agent

   • C

     Place the Lambda function in the same VPC subnet as the agent so the two can communicate

   • D

     Attach AWSLambdaBasicExecutionRole to the agent service role to enable invocation of functions

   Show answer & explanation

   Correct answer: B — Add a resource-based policy on the Lambda function that lets bedrock.amazonaws.com invoke it for that agent

   Bedrock invokes the function as a service principal, so the function needs a resource-based policy permitting bedrock.amazonaws.com, ideally limited to the agent's ARN. The execution role governs what the function can call, not who can call it. Agents are not placed in subnets, and the basic execution role grants log-writing permissions for Lambda, not invocation rights.

7. Question 8ML Model and Foundation Model (FM) DevelopmentMedium

   A team deploys a Neo-compiled PyTorch model with a SageMaker AI Neo Inference Optimized Container and writes an inference.py entry point. They are unsure what the script must define. Which statement about model_fn is correct?

   • A

     It is mandatory, and the endpoint fails to start whenever the inference script omits this function

   • B

     It converts the incoming request payload into a numpy array, which is done before the prediction call

   • C

     It is optional, since a default loads the Neo-compiled model unless extra loading steps are needed

   • D

     It replaces predict_fn and output_fn, so the script needs no other handler functions beyond this one

   Show answer & explanation

   Correct answer: C — It is optional, since a default loads the Neo-compiled model unless extra loading steps are needed

   In the Neo Inference Optimized Containers for PyTorch and MXNet, model_fn loads the model and is marked optional, because a default model_fn already loads the model compiled with SageMaker Neo. A team needs to define it when the model requires extra steps to load. Converting the request payload into a numpy array is the job of input_fn, and transform_fn, not model_fn, is the function that can combine input_fn, predict_fn, and output_fn.

8. Question 9Data Preparation for ML and AIHard

   A data team owns a Glue Data Catalog table with hundreds of columns and no documented quality rules. A nightly Glue ETL job must publish just the rows that pass checks to a curated S3 prefix while sending failing rows to a quarantine prefix for review. Which approach meets these requirements with the least manual rule authoring?

   • A

     Use a SageMaker Model Monitor data quality baseline on the table and route violations to quarantine with EventBridge

   • B

     Run a DataBrew profile job and add a Lambda function that deletes failing rows from the curated prefix after each load

   • C

     Run a crawler to infer the schema and fail the job whenever the crawler detects a new column in the nightly files

   • D

     Run a Glue Data Quality rule recommendation on the table, then split passing and failing rows using row-level outcomes

   Show answer & explanation

   Correct answer: D — Run a Glue Data Quality rule recommendation on the table, then split passing and failing rows using row-level outcomes

   Glue Data Quality can generate a recommended DQDL ruleset from the table data, which removes most manual rule writing. The Evaluate Data Quality transform can emit row-level outcomes, so passing and failing rows are written to separate prefixes in the same job. Model Monitor baselines apply to SageMaker endpoint traffic, not catalog tables. The DataBrew and Lambda approach publishes bad rows first and deletes them later, and a crawler schema check does not evaluate row values.