15 Free AWS Certified Machine Learning Engineer – Associate (MLA-C02) Practice Questions with Answers (2026)

Splitting first keeps the test set a true sample of real-world data. A stratified split keeps the 0.5% ratio in both sets, and oversampling only the training set fixes the imbalance without letting synthetic rows leak into evaluation.

• A is wrong because synthetic rows created from the full dataset end up in the test set, which inflates the results.
• B is wrong because it throws away the minority class you are trying to learn.
• D is wrong because a model that always predicts "not fraud" scores 99.5% accuracy, so accuracy hides the problem.

Amazon Data Firehose can convert incoming JSON records to Parquet (or ORC) before delivering them to S3, using a table schema from the Glue Data Catalog. It is fully managed and columnar Parquet reduces the data Athena scans.

• B is wrong because an EC2 script adds a server to patch and run, and the data is only converted once a day.
• C is wrong because SQS is a message queue, not storage for analytics, and it does not produce Parquet.
• D is wrong because a relational database adds cost and operations, and CSV is a row format that is slower to query than Parquet.

Amazon Comprehend can run asynchronous PII detection jobs over documents in S3 and write redacted copies, replacing detected entities such as names, phone numbers and addresses. That gives you a cleaned dataset without writing redaction logic.

• A is wrong because Macie discovers and reports sensitive data in S3; it does not produce redacted copies.
• B is wrong because encryption protects data at rest, but the model would still be trained on the PII once decrypted.
• C is wrong because Textract extracts text and structure from scanned documents; it does not detect or redact PII.

The team wants to change the model's behaviour (format and labels) and has labelled pairs. That is exactly what supervised fine-tuning does, and it removes the need for long few-shot prompts.

• A is wrong because continued pre-training uses large amounts of unlabelled text to teach domain language, not a specific task format.
• B is wrong because retrieval adds facts to the prompt; it does not reliably teach the model a consistent output behaviour.
• D is wrong because higher temperature makes outputs more random, which makes format drift worse.

ROUGE (Recall-Oriented Understudy for Gisting Evaluation) measures overlap between a generated summary and reference summaries, with a focus on recall. It is the standard automatic metric for summarisation.

• A is wrong because BLEU is precision-focused and designed for machine translation.
• C is wrong because perplexity measures how well a model predicts text, not how a summary compares with a reference.
• D is wrong because AUC evaluates binary classifiers, not generated text.

Falling training loss with rising validation loss is overfitting. Early stopping halts training when the validation metric stops improving, and regularisation limits how closely the model fits noise in the training data.

• A is wrong because more rounds makes the model fit the training data even more closely.
• B is wrong because deeper trees increase model complexity, which also increases overfitting.
• C is wrong because without a validation set you cannot detect overfitting at all.

Amazon Textract analyses scanned documents and, with the forms feature, returns key-value pairs such as a field label and its value. It is managed and needs no training data.

• B is wrong because Comprehend works on plain text; it does not understand the layout of a scanned form.
• C is wrong because Rekognition detects text in images but does not link labels to their values.
• D is wrong because there is no labelled data, and building a custom model is far more effort than a managed service that already does this.

Asynchronous inference queues requests, supports large payloads and processing times of many minutes, and can scale down to zero instances when there is nothing to process. That matches large, slow, bursty requests.

• A is wrong because real-time endpoints are built for small, fast requests, and fixed instances cost money during idle periods.
• C is wrong because serverless inference has small payload limits and short timeouts, which these requests exceed.
• D is wrong because users would wait up to a week for results.

AgentCore Runtime is serverless hosting for agents built with any framework, including Strands. Each session runs in its own isolated microVM, and sessions can run for hours, which covers the two-hour requirement.

• A is wrong because a Lambda invocation has a 15-minute maximum, well short of two hours.
• B is wrong because SageMaker AI endpoints host models for inference, not long-running stateful agent sessions.
• D is wrong because a knowledge base stores and retrieves documents; it does not run agent code.

A knowledge base sync only processes documents that were added, changed or deleted since the last sync, so a daily scheduled ingestion job keeps answers current at low cost.

• B is wrong because recreating the knowledge base re-embeds every document each night, which costs more and adds downtime.
• C is wrong because returning more results does nothing if the index still holds old content.
• D is wrong because daily fine-tuning is slow and expensive, and it does not give traceable answers from the source documents.

Model Registry keeps versioned model packages with an approval status and lineage for audits. When a reviewer changes a version to Approved, SageMaker emits an event that an EventBridge rule can use to start the deployment pipeline.

• A is wrong because it deploys without any approval step.
• B is wrong because manual deployment is not automated and tags do not give an approval workflow or audit history.
• C is wrong because Git is not built to store large model artifacts and has no model approval status.

Denied topics block conversations about a subject you define, such as personal investment advice. Sensitive information filters detect PII such as email addresses, plus custom patterns such as account numbers, and block or mask them.

• A is wrong because contextual grounding checks whether a response is supported by the source material; it does not block a subject or PII.
• C is wrong because the prompt attack filter detects jailbreaks and prompt injection, not a business topic.
• E is wrong because image filters handle harmful image content, which is not the requirement.

Model Monitor compares captured live requests against a baseline built from the training data and reports violations, such as distribution drift, as CloudWatch metrics you can alarm on.

• A is wrong because a bigger instance changes latency, not accuracy.
• B is wrong because CloudTrail records API calls for auditing; it does not analyse the feature values in requests.
• D is wrong because blind daily retraining adds cost and still gives no visibility into what changed.

A task role gives the container temporary, automatically rotated credentials, and a least-privilege policy can restrict invocation to specific model ARNs. Nothing long-lived is stored.

• B is wrong because AWS recommends long-term Bedrock API keys only for exploration, and an environment variable exposes the key.
• C is wrong because short-term keys expire within 12 hours, so a weekly manual key would break, and the process is manual.
• D is wrong because IAM user access keys are long-lived and baking them into an image exposes them to anyone with the image.