Skip to main content
CertRampLearning

AWS · Free practice questions

AWS MLA-C02 Practice Questions

15 original MLA-C02 practice questions across the four content domains of the AWS exam guide, weighted roughly like the real exam. They cover both traditional ML on Amazon SageMaker AI and generative AI on Amazon Bedrock. Each one comes with the answer, a short explanation and the reason every other option is wrong.

Written for the mLA-C02 exam guide (beta exam; successor to MLA-C01) · Updated October 27, 2026

These are original practice questions written by CertRamp to the official AWS exam outline. They are not real exam questions and do not come from any exam. Try each question first, then open the answer to see the explanation for every option.

Domain 1

Data Preparation for ML and AI

1.1 Collect and store dataFoundationDomain: Data Preparation for ML and AI

Question 1. A team stores several terabytes of training data in Amazon S3 as CSV files with 300 columns. Most Amazon Athena queries and processing jobs read only 10–15 columns and filter on a few of them. Which change best reduces the amount of data scanned?

  1. Option A: Convert the files to JSON Lines
  2. Option B: Combine all CSV files into one large file
  3. Option C: Convert the data to Apache Parquet
  4. Option D: Compress the CSV files with gzip
Show answer and explanation

Correct answer: C. Convert the data to Apache Parquet

The exam guide asks you to choose data formats such as Parquet, JSON, CSV and ORC based on access patterns. When queries touch a few columns of wide tables, a columnar format like Parquet or ORC is the standard choice.

  • A (incorrect)JSON is a row-based text format. Every record still has to be read in full, and the field names repeated in each record add size.
  • B (incorrect)File consolidation can reduce overhead from many small files, but a single CSV file is still row-based and is scanned in full.
  • C (correct)Correct. Parquet is a columnar format with per-column compression, and block statistics allow engines such as Athena to skip data they do not need.
  • D (incorrect)Compression reduces storage, but a gzip CSV file still has to be decompressed and read row by row, including the unused columns.

Reference: Amazon Athena: Use columnar storage formats (opens in a new tab)

1.2 Perform data transformation, feature engineering, and pre-processingIntermediateDomain: Data Preparation for ML and AI

Question 2. A fraud model needs the latest customer features at inference time with low latency, and the data science team needs the full feature history to build training datasets. The team wants one managed service for both. What should you use?

  1. Option A: Amazon SageMaker Feature Store with both the online store and the offline store enabled
  2. Option B: Amazon SageMaker Feature Store with only the offline store enabled
  3. Option C: AWS Glue DataBrew recipes
  4. Option D: An Amazon S3 bucket queried with Amazon Athena at inference time
Show answer and explanation

Correct answer: A. Amazon SageMaker Feature Store with both the online store and the offline store enabled

Feature Store has two stores for two jobs: online for low-latency lookup of the latest values, offline for history used in training and batch inference. Ingesting into one feature group with both stores enabled helps keep training and serving features consistent.

  • A (correct)Correct. The online store serves the latest record for real-time lookup, the offline store keeps the full history in Amazon S3 for training, and by default each ingested record is written to both stores.
  • B (incorrect)The offline store is for cases where sub-second reads are not needed, such as training and batch inference. It does not meet the low-latency requirement.
  • C (incorrect)DataBrew prepares and cleans data. It does not serve features to models at inference time.
  • D (incorrect)Athena queries on S3 are not designed for low-latency lookups during real-time inference.

Reference: Amazon SageMaker AI: Feature Store concepts (opens in a new tab)

1.2 Perform data transformation, feature engineering, and pre-processingAdvancedDomain: Data Preparation for ML and AI

Question 3. You are building an Amazon Bedrock knowledge base over long technical manuals. Retrieval must match on small, precise passages, but the model should then receive the wider surrounding section as context. Which chunking strategy fits best?

  1. Option A: No chunking
  2. Option B: Fixed-size chunking with a large overlap
  3. Option C: Default chunking
  4. Option D: Hierarchical chunking
Show answer and explanation

Correct answer: D. Hierarchical chunking

Preparing documents for RAG, including chunking strategies, is a named skill in the exam guide. Hierarchical chunking separates the unit used for matching (child) from the unit passed to the model (parent).

  • A (incorrect)Each document becomes a single chunk, so retrieval cannot match precise passages within long manuals.
  • B (incorrect)Overlap reduces the risk of cutting information at chunk boundaries, but every chunk is still one size, so you trade precision against context.
  • C (incorrect)Default chunking splits text into chunks of roughly 300 tokens. It does not return a larger parent section after matching.
  • D (correct)Correct. Hierarchical chunking creates parent and child chunks. Retrieval matches the small child chunks and replaces them with their parent chunks to give the model more context.

Reference: Amazon Bedrock: How content chunking works for knowledge bases (opens in a new tab)

1.3 Validate data quality and manage biasIntermediateDomain: Data Preparation for ML and AI

Question 4. Before training a loan approval model, you must check whether one age group is heavily under-represented in the training data compared with another. No model exists yet. Which approach is most appropriate?

  1. Option A: Calculate SHAP values with SageMaker Clarify
  2. Option B: Run a SageMaker Clarify pre-training bias analysis and review the Class Imbalance (CI) metric for the age facet
  3. Option C: Deploy the model and compare predictions by age group
  4. Option D: Run AWS Glue Data Quality rules for completeness
Show answer and explanation

Correct answer: B. Run a SageMaker Clarify pre-training bias analysis and review the Class Imbalance (CI) metric for the age facet

Clarify's pre-training metrics are model-agnostic and run on the data alone. A CI value near −1 or 1 signals strong imbalance between facets, which you may want to fix by rebalancing the sample before training.

  • A (incorrect)SHAP values explain how features influence a trained model's predictions, so they require a model.
  • B (correct)Correct. Pre-training bias metrics are computed on the raw dataset, and CI measures whether one facet has fewer samples than another.
  • C (incorrect)This measures post-training bias and requires a trained model, which does not exist yet.
  • D (incorrect)Completeness checks find missing values. They do not measure the representation of one group relative to another.

Reference: Amazon SageMaker AI: Class Imbalance (CI) (opens in a new tab)

Domain 2

ML Model and Foundation Model (FM) Development

2.1 Choose appropriate modeling approaches for ML and AI solutionsFoundationDomain: ML Model and Foundation Model (FM) Development

Question 5. An insurer receives thousands of scanned claim forms each day and needs to extract form fields and tables into a structured format. The team has no labelled data and wants the least development effort. What should you use?

  1. Option A: Amazon Textract
  2. Option B: Amazon Comprehend
  3. Option C: Amazon Rekognition
  4. Option D: A custom computer vision model trained with SageMaker AI
Show answer and explanation

Correct answer: A. Amazon Textract

The exam guide expects you to apply AWS AI services to specific business problems and to weigh managed services against custom models. When a managed service already does the job, it is usually the lowest-effort answer.

  • A (correct)Correct. Textract is a managed AI service that detects typed and handwritten text and extracts forms and tables from documents, without training a model.
  • B (incorrect)Comprehend analyses text, for example for entities or sentiment. It does not extract form fields and tables from scanned images.
  • C (incorrect)Rekognition analyses images and video, such as objects and faces. It is not designed for structured document extraction.
  • D (incorrect)This is possible but requires labelled data and much more effort than a managed service built for the task.

Reference: Amazon Textract: What is Amazon Textract? (opens in a new tab)

2.1 Choose appropriate modeling approaches for ML and AI solutionsIntermediateDomain: ML Model and Foundation Model (FM) Development

Question 6. A training job runs for about 10 hours each week. It is not time-critical and the team wants to reduce training cost significantly. The training script can save its state. What should you configure?

  1. Option A: A larger instance type so the job finishes sooner
  2. Option B: A real-time endpoint that stays running between training runs
  3. Option C: Managed spot training with checkpointing to Amazon S3
  4. Option D: Managed spot training without checkpointing and with MaxWaitTimeInSeconds equal to MaxRuntimeInSeconds
Show answer and explanation

Correct answer: C. Managed spot training with checkpointing to Amazon S3

Spot training trades guaranteed capacity for lower cost, which suits jobs that are not time-critical. Managed spot training does not fall back to on-demand instances, so checkpointing and a sensible wait time are essential.

  • A (incorrect)A larger instance may shorten the job, but it does not necessarily reduce cost, and it ignores the cheapest compute option available.
  • B (incorrect)Endpoints serve inference. Keeping one running adds cost and does not train the model.
  • C (correct)Correct. Managed spot training uses Spot capacity, which AWS says can cut training costs by up to 90% compared with on-demand. Checkpoints let an interrupted job resume instead of restarting.
  • D (incorrect)Without checkpoints, every Spot interruption restarts the 10-hour job from the beginning, and a wait time equal to the runtime leaves no time to wait for Spot capacity or to recover from interruptions.

Reference: Amazon SageMaker AI: Managed Spot Training (opens in a new tab)

2.2 Train, fine-tune, and customize models for ML and AI solutionsFoundationDomain: ML Model and Foundation Model (FM) Development

Question 7. During training, the training loss keeps falling, but after epoch 12 the validation loss starts to rise steadily. What is happening, and what is a suitable response?

  1. Option A: The model is underfitting; add more epochs
  2. Option B: The model is overfitting; use early stopping or add regularisation
  3. Option C: The learning rate is too low; increase it sharply
  4. Option D: The validation set is too large; remove most of it
Show answer and explanation

Correct answer: B. The model is overfitting; use early stopping or add regularisation

Preventing overfitting and underfitting is a named skill, and early stopping is listed as a training time reduction technique. Watch the gap between training and validation metrics, not the training loss alone.

  • A (incorrect)Underfitting shows as high loss on both training and validation data. More epochs would make the observed problem worse.
  • B (correct)Correct. A falling training loss with a rising validation loss means the model is memorising the training data. Early stopping, regularisation, dropout or more data help.
  • C (incorrect)A learning rate that is too low slows both losses down. It does not explain diverging training and validation losses.
  • D (incorrect)Shrinking the validation set only hides the problem and makes the evaluation less reliable.

Reference: AWS: MLA-C02 exam guide, Content Domain 2 (opens in a new tab)

2.3 Analyze and evaluate the performance of ML and AI systemsIntermediateDomain: ML Model and Foundation Model (FM) Development

Question 8. A team wants to compare two foundation models on 500 of its own prompts, scoring each response for quality with explanations. It does not have reference answers and cannot staff a human review team. Which approach fits best?

  1. Option A: Calculate ROUGE scores
  2. Option B: Amazon Bedrock model evaluation jobs with a human work team
  3. Option C: Compare the CloudWatch invocation latency of both models
  4. Option D: Amazon Bedrock model evaluation jobs that use a model as a judge, one for each model
Show answer and explanation

Correct answer: D. Amazon Bedrock model evaluation jobs that use a model as a judge, one for each model

The exam guide names LLM-as-a-judge frameworks, human-in-the-loop evaluation and NLP metrics such as BLEU and ROUGE. Pick the method that fits what you have: references, reviewers or neither.

  • A (incorrect)ROUGE compares generated text with reference text. Without reference answers it cannot be calculated.
  • B (incorrect)Human evaluation is valuable, but the team cannot staff reviewers.
  • C (incorrect)Latency is a useful operational metric, but it says nothing about response quality.
  • D (correct)Correct. In LLM-as-a-judge evaluation, an evaluator model scores each prompt and response pair against the selected metrics and explains its scores. Each job evaluates one generator model, so you run one job per model and compare the results.

Reference: Amazon Bedrock: Model evaluation jobs that use a model as a judge (opens in a new tab)

Domain 3

Deployment and Orchestration of ML and AI Workflows

3.1 Manage deployment infrastructure for ML and AI model typesIntermediateDomain: Deployment and Orchestration of ML and AI Workflows

Question 9. A model processes uploaded audio files of up to 500 MB. Each request takes 10–20 minutes, requests arrive irregularly with long idle periods, and users can wait for the result. Which SageMaker AI inference option fits best?

  1. Option A: Asynchronous inference
  2. Option B: Real-time inference
  3. Option C: Serverless inference
  4. Option D: Batch transform scheduled once a month
Show answer and explanation

Correct answer: A. Asynchronous inference

Choosing between real-time, serverless, asynchronous and batch inference is a core deployment decision. Large payloads, long processing and near-real-time expectations point to asynchronous inference.

  • A (correct)Correct. Asynchronous inference queues requests, supports payloads up to 1 GB stored in Amazon S3 and processing times up to one hour, and can scale to zero instances when idle.
  • B (incorrect)Real-time endpoints are designed for low-latency, interactive requests, not for large payloads that take many minutes to process.
  • C (incorrect)Serverless inference suits intermittent traffic with small, quick requests. It is not designed for 500 MB payloads and 20-minute processing.
  • D (incorrect)Batch transform processes whole datasets offline. A monthly schedule would leave users waiting far too long.

Reference: Amazon SageMaker AI: Asynchronous inference (opens in a new tab)

3.1 Manage deployment infrastructure for ML and AI model typesAdvancedDomain: Deployment and Orchestration of ML and AI Workflows

Question 10. A team fine-tuned a Llama 3.1 model outside AWS and wants to call it through the same Amazon Bedrock APIs it uses for other models, without managing inference instances. What should it do?

  1. Option A: Upload the model to an Amazon Bedrock knowledge base
  2. Option B: Package the model as a Bedrock agent action group
  3. Option C: Store the weights in Hugging Face format (including .safetensors files) in Amazon S3 and create an Amazon Bedrock Custom Model Import job
  4. Option D: Convert the model with SageMaker Neo and deploy it to an edge device
Show answer and explanation

Correct answer: C. Store the weights in Hugging Face format (including .safetensors files) in Amazon S3 and create an Amazon Bedrock Custom Model Import job

The exam guide names Amazon Bedrock Custom Model Import for deploying models built outside AWS. Check the supported architectures and file requirements, and remember that SageMaker AI is the alternative when you need full control of hosting.

  • A (incorrect)Knowledge bases store and retrieve documents for RAG. They do not host models.
  • B (incorrect)Action groups define the actions an agent can call. They do not host a foundation model.
  • C (correct)Correct. Custom Model Import imports supported architectures such as Llama from Amazon S3 and lets you invoke the model with InvokeModel, with on-demand throughput.
  • D (incorrect)Neo optimises models for edge targets. That does not make the model callable through Amazon Bedrock.

Reference: Amazon Bedrock: Custom Model Import (opens in a new tab)

3.2 Provision and configure resources for ML and AI workloads based on existing architecture and requirementsIntermediateDomain: Deployment and Orchestration of ML and AI Workflows

Question 11. A travel booking agent runs on Amazon Bedrock AgentCore. It must remember the context within a conversation and also recall user preferences, such as a preferred seat, in later sessions, without the team building its own storage layer. What should you use?

  1. Option A: A larger context window on the model
  2. Option B: AgentCore Memory with short-term and long-term memory
  3. Option C: Amazon Bedrock Prompt Management
  4. Option D: CloudWatch Logs Insights queries on past conversations
Show answer and explanation

Correct answer: B. AgentCore Memory with short-term and long-term memory

Implementing agent state management systems is a new skill in MLA-C02. AgentCore Memory provides managed short-term and long-term memory, so agents are not stateless between interactions.

  • A (incorrect)A larger context window helps within one request, but it does not persist preferences across separate sessions.
  • B (correct)Correct. Short-term memory keeps turn-by-turn context within a session, and long-term memory extracts and stores insights such as user preferences across sessions.
  • C (incorrect)Prompt Management stores and versions prompts. It does not keep per-user conversation state.
  • D (incorrect)Logs support troubleshooting and analysis. They are not a memory store the agent uses at runtime.

Reference: Amazon Bedrock AgentCore: Memory (opens in a new tab)

3.3 Implement automated orchestration and continuous integration and continuous delivery (CI/CD) pipelines for MLOps and AI workloadsIntermediateDomain: Deployment and Orchestration of ML and AI Workflows

Question 12. A SageMaker AI pipeline retrains a model weekly. A reviewer must sign off each new model version before deployment, and the decision must be recorded for audits. Deployment should then start automatically. Which design fits best?

  1. Option A: Deploy every new model to production and roll back if complaints arrive
  2. Option B: Email the model artefact location to the reviewer and deploy manually after a reply
  3. Option C: Store each model in a new S3 prefix and let the reviewer rename the folder
  4. Option D: Register each version in SageMaker Model Registry with the status PendingManualApproval, and let a change to Approved start the CI/CD deployment
Show answer and explanation

Correct answer: D. Register each version in SageMaker Model Registry with the status PendingManualApproval, and let a change to Approved start the CI/CD deployment

The exam guide lists managing model versions for repeatability and audits, for example with SageMaker Model Registry. The approval status becomes the gate between training and deployment.

  • A (incorrect)This skips the required sign-off and leaves no approval record.
  • B (incorrect)This is manual, error-prone and leaves no structured record linked to the model version.
  • C (incorrect)S3 prefixes do not provide versioned approval status or deployment integration.
  • D (correct)Correct. Model Registry tracks versions with an approval status, and with a SageMaker AI project template, approving a version initiates CI/CD deployment.

Reference: Amazon SageMaker AI: Update the approval status of a model (opens in a new tab)

Domain 4

Operating, Monitoring, and Securing ML and AI Solutions

4.1 Monitor ML and AI model inference and performanceIntermediateDomain: Operating, Monitoring, and Securing ML and AI Solutions

Question 13. You want to send 10% of live traffic to a new model version on an existing SageMaker AI real-time endpoint, compare it with the current model, and then gradually shift traffic without redeploying the endpoint. What should you do?

  1. Option A: Add the new model as a second production variant with a lower weight, compare per-variant metrics, then adjust weights with UpdateEndpointWeightsAndCapacities
  2. Option B: Deploy the new model to a separate endpoint and switch DNS once a week
  3. Option C: Run both models in a batch transform job each night
  4. Option D: Set TargetVariant on every request to the new model
Show answer and explanation

Correct answer: A. Add the new model as a second production variant with a lower weight, compare per-variant metrics, then adjust weights with UpdateEndpointWeightsAndCapacities

Monitoring model performance in production with A/B testing is a named skill. Weighted production variants give you a controlled split, and per-variant CloudWatch metrics show how each version behaves.

  • A (correct)Correct. Production variant weights control the traffic split, and UpdateEndpointWeightsAndCapacities changes them without updating the endpoint.
  • B (incorrect)This creates an all-or-nothing switch and gives no controlled, weighted split on one endpoint.
  • C (incorrect)Batch transform does not test the model on live traffic.
  • D (incorrect)TargetVariant forces each request to one variant. Setting it on every request would send all traffic to the new model, not 10%.

Reference: Amazon SageMaker AI: Test models with production variants (opens in a new tab)

4.2 Optimize and manage ML and AI infrastructure costs and performanceIntermediateDomain: Operating, Monitoring, and Securing ML and AI Solutions

Question 14. A small CPU-based model behind an internal tool receives a few hundred requests per day, mostly during office hours, and is idle overnight and at weekends. An always-on ml.m5.large endpoint is costing more than the tool is worth. Occasional cold-start latency is acceptable. What should you do?

  1. Option A: Move the model to a GPU instance
  2. Option B: Deploy the model with SageMaker Serverless Inference
  3. Option C: Add a second instance to the existing endpoint for availability
  4. Option D: Use Provisioned Concurrency for 24 hours a day
Show answer and explanation

Correct answer: B. Deploy the model with SageMaker Serverless Inference

Matching capacity to the traffic pattern is a core cost skill. Intermittent, unpredictable traffic with tolerance for cold starts points to serverless inference; predictable bursts may justify Provisioned Concurrency.

  • A (incorrect)A GPU instance costs more and the CPU-based model does not need it.
  • B (correct)Correct. Serverless Inference is pay-per-use, scales the endpoint down to zero when there are no requests and suits intermittent traffic where cold starts are acceptable.
  • C (incorrect)This doubles the cost of an endpoint that is already underused.
  • D (incorrect)Provisioned Concurrency keeps capacity warm and is billed for it, which removes most of the saving for a mostly idle endpoint.

Reference: Amazon SageMaker AI: Serverless Inference (opens in a new tab)

4.3 Secure ML and AI workloads and model endpointsAdvancedDomain: Operating, Monitoring, and Securing ML and AI Solutions

Question 15. A customer support assistant built on Amazon Bedrock must never show email addresses or phone numbers in its responses. Responses should still be returned, with the personal data replaced by a placeholder. What should you configure?

  1. Option A: A denied topic for personal data in Amazon Bedrock Guardrails
  2. Option B: Encrypt the model invocation logs with an AWS KMS key
  3. Option C: An Amazon Bedrock guardrail with a sensitive information filter that masks EMAIL and PHONE in outputs
  4. Option D: An IAM policy that denies bedrock:InvokeModel for responses containing PII
Show answer and explanation

Correct answer: C. An Amazon Bedrock guardrail with a sensitive information filter that masks EMAIL and PHONE in outputs

The exam guide names Amazon Bedrock Guardrails for safeguards and sensitive data protection. Choose mask when the response should still be delivered, and block when any detection should stop the response.

  • A (incorrect)Denied topics block whole subjects of conversation. They do not replace individual pieces of personal data inside an otherwise valid response.
  • B (incorrect)Encryption protects stored logs. It does not change what the model returns to the user.
  • C (correct)Correct. Sensitive information filters can block or mask PII. Masking replaces the detected value with its type, such as {EMAIL}, and can be set separately for inputs and outputs.
  • D (incorrect)IAM controls who may call an API. It cannot inspect or alter the content of a model response.

Reference: Amazon Bedrock: Remove PII from conversations by using sensitive information filters (opens in a new tab)

Want more practice?

15 questions are a sample. The exam is 85 questions in the beta exam; the exam guide specifies 50 scored and 15 unscored questions items.

The full CertRamp AWS MLA-C02 practice exams have 6 tests that grow with you, from a diagnostic start to an exam-level Test 4 and a challenge set beyond the exam, with an explanation for every option.

Who writes the exams

Written by a practitioner, not a question mill

Joshua Ravnjak, founder of CertRamp Learning

Joshua Ravnjak

Senior Consultant, digital transformation and AI

Founder, CertRamp Learning

921 learners and 61 reviews on Udemy (as of November 3, 2026)

I have more than 12 years of professional experience and work as a senior consultant in an IT management consultancy, focusing on digital transformation, AI and requirements engineering in the insurance industry. I have earned many certifications myself, so I know what exam preparation feels like. That is why every CertRamp practice exam follows the same progression: an easy start to see where you stand, then real exam level, then beyond it. You walk into the exam with a buffer and a good feeling.

Selected certifications

  • IREB® CPRE Foundation Level
  • ITIL® 4 Foundation
  • Professional Scrum Master I (PSM I)
  • Professional Scrum Product Owner I (PSPO I)
  • Certified SAFe® 6 Scrum Master
  • Certified AI-Empowered SAFe® 6 Product Owner/Product Manager

Plus further certifications from Microsoft and Anthropic (Claude).

ITIL® is a registered trademark of PeopleCert. IREB® is a registered trademark of IREB e.V. SAFe® is a registered trademark of Scaled Agile, Inc. PSM and PSPO are trademarks of Scrum.org.

Also in print

Cover: CPRE Foundation Level Prüfungsvorbereitung 2026 by Joshua Ravnjak

CPRE Foundation Level Prüfungsvorbereitung 2026

6 Übungsprüfungen mit 270 Fragen und ausführlichen Erklärungen

Paperback, German, October 2026

My CPRE Foundation Level exam-prep book (in German) follows the same progressive approach.

View on Amazon (opens in a new tab)