AWS MLA-C02 Sample Questions and Exam Guide (Beta)
10 original MLA-C02 practice questions for AWS Certified Machine Learning Engineer Associate, plus beta exam facts, the four domains and what changed from MLA-C01.
In this article
- MLA-C02 exam facts at a glance
- Domain 1: Data Preparation for ML and AI
- Domain 2: ML Model and Foundation Model (FM) Development
- Domain 3: Deployment and Orchestration of ML and AI Workflows
- Domain 4: Operating, Monitoring, and Securing ML and AI Solutions
- What changed from MLA-C01 to MLA-C02
- How to score yourself
- How these questions were written
- Practice more MLA-C02 questions
MLA-C02 is the updated AWS Certified Machine Learning Engineer - Associate exam, and AWS is running it as a beta right now. Below are 10 original practice questions across its four domains, with answers and explanations, followed by the beta facts and what changed from MLA-C01.
Key takeaways
- MLA-C02 tests ML engineering on AWS: data pipelines, training, deployment, monitoring and security, for both traditional ML and foundation models.
- Three of the four domains are weighted 24% each; data preparation is 28%.
- It is an associate exam. AWS recommends at least 1 year of hands-on experience with SageMaker AI and Bedrock.
MLA-C02 exam facts at a glance
Facts come from the AWS certification page and the MLA-C02 exam guide, checked on 2026-11-02. More detail and a study plan are on our MLA-C02 exam prep page. For a week-by-week schedule, see how to study for MLA-C02.
| Fact | Official value |
|---|---|
| Status | Beta; general availability date to be announced |
| Beta format | 85 questions, 170 minutes, multiple choice and multiple response |
| Beta price | 75 USD |
| Language | English only during the beta |
| Passing score (exam guide) | 720 on a scaled score of 100-1,000 |
| Valid for | 3 years |
| Domain | Weight | Questions below |
|---|---|---|
| 1. Data Preparation for ML and AI | 28% | 1-3 |
| 2. ML Model and Foundation Model (FM) Development | 24% | 4-5 |
| 3. Deployment and Orchestration of ML and AI Workflows | 24% | 6-7 |
| 4. Operating, Monitoring, and Securing ML and AI Solutions | 24% | 8-10 |
Domain 1: Data Preparation for ML and AI
Question 1
A team trains a model on a dataset where 98% of examples belong to one class. The model scores 98% accuracy but never detects the rare class. What should they do first?
- A. Report the 98% accuracy as the result
- B. Address the class imbalance, for example by resampling, and evaluate with precision and recall
- C. Remove the rare class from the data
- D. Reduce the number of features to one
Answer: B. Accuracy hides failure on imbalanced data. Rebalancing (oversampling the rare class or undersampling the common one) and judging by precision, recall or F1 gives a truthful picture.
- A is wrong because a model that always predicts the majority class also scores 98%.
- C is wrong because the rare class is usually the one that matters.
- D is wrong because fewer features do not fix imbalance.
Question 2
A data scientist wants to reuse the same engineered features for training and for low-latency online predictions, without re-implementing them twice. Which approach fits?
- A. Store features in a central feature store such as Amazon SageMaker Feature Store
- B. Email feature definitions to the application team
- C. Recompute features by hand in each notebook
- D. Store features only in local CSV files
Answer: A. A feature store keeps one governed definition of each feature and serves it for both offline training and online inference, which prevents training-serving skew.
- B, C and D are wrong because they duplicate logic, which is how training and serving features drift apart.
Question 3
A team builds a RAG application and must prepare company documents so a model can retrieve relevant passages. Which step is part of that preparation?
- A. Split the documents into chunks and create embeddings stored in a vector store
- B. Delete all documents longer than one page
- C. Convert every document into an image
- D. Train a new model from scratch for every document
Answer: A. RAG preparation chunks documents, turns the chunks into embeddings and stores them in a vector store so similar passages can be retrieved at query time. See what is RAG.
- B, C and D are wrong because they destroy content, lose text or are needlessly expensive.
Domain 2: ML Model and Foundation Model (FM) Development
Question 4
A team has a small labeled dataset and needs an image classifier quickly. Training from scratch gives poor results. What is the most effective approach?
- A. Use transfer learning from a pretrained model and fine-tune it on the labeled data
- B. Collect no more data and train longer
- C. Use a linear regression model
- D. Randomly assign labels
Answer: A. Transfer learning reuses features a model learned on a large dataset, so a small labeled set is enough to adapt it.
- B is wrong because training longer on too little data tends to overfit.
- C is wrong because regression predicts numbers, not image classes.
- D is wrong because random labels teach nothing.
Question 5
A company wants to compare two foundation models for a summarization task. Which is the most appropriate evaluation approach?
- A. Run both models on the same representative test prompts and compare results with defined metrics or human or model-based judging
- B. Pick the model with the longer name
- C. Evaluate only on the training prompts of one model
- D. Skip evaluation because both are foundation models
Answer: A. Evaluate on a fixed, representative set with clear criteria. Automated metrics, human reviewers or a judge model can all be used, ideally together.
- B, C and D are wrong because they are not valid evaluation or are biased toward one model.
Domain 3: Deployment and Orchestration of ML and AI Workflows
Question 6
A model receives unpredictable traffic with long idle periods, and requests can tolerate a short cold start. Which deployment option minimizes cost?
- A. A serverless inference endpoint
- B. A large always-on real-time endpoint with many instances
- C. A dedicated GPU instance running 24/7 for no traffic
- D. Retraining the model for every request
Answer: A. Serverless inference scales to zero when idle and charges for use, which suits intermittent traffic that can accept cold starts.
- B and C are wrong because they pay for idle capacity.
- D is wrong because inference does not require retraining.
Question 7
A team wants every change to its training code to automatically trigger tests, a training pipeline and a model-registry step. What should they build?
- A. A CI/CD pipeline integrated with an ML pipeline and a model registry
- B. A weekly manual checklist
- C. A shared spreadsheet of model versions
- D. A single notebook that everyone edits
Answer: A. MLOps automates build, test, train, register and deploy steps so releases are repeatable and auditable.
- B, C and D are wrong because they are manual and error-prone.
Domain 4: Operating, Monitoring, and Securing ML and AI Solutions
Question 8
A deployed model's accuracy slowly drops because customer behavior has changed since training. What is this called, and what should be set up?
- A. Data drift; monitor input and prediction distributions and trigger retraining
- B. Overfitting; add more layers
- C. Underfitting; reduce the data
- D. Latency; add more storage
Answer: A. Production data shifting away from the training data is drift. Monitoring (for example with SageMaker Model Monitor) detects it so you can retrain.
- B and C are wrong because they describe training-time fit problems.
- D is wrong because latency is about speed, not accuracy.
Question 9
A training job in SageMaker AI must read data from an S3 bucket. Which approach follows least-privilege security?
- A. Attach an IAM role to the job that grants access only to that bucket and prefix
- B. Embed long-lived access keys in the training code
- C. Make the bucket public
- D. Give the job administrator access
Answer: A. An execution role with narrowly scoped permissions gives temporary credentials and limits the damage of any mistake.
- B is wrong because hard-coded keys leak easily.
- C and D are wrong because they grant far more access than needed.
Question 10
A company needs traffic between its VPC and its model endpoints to stay off the public internet. Which option helps?
- A. Use VPC interface endpoints (AWS PrivateLink) for the service
- B. Put the endpoint's URL on a public website
- C. Disable encryption
- D. Open all security group ports
Answer: A. Interface VPC endpoints keep traffic on the AWS network, combined with security groups and encryption.
- B, C and D are wrong because they increase exposure.
What changed from MLA-C01 to MLA-C02
AWS publishes a comparison of MLA-C01 and MLA-C02. In short, the updated exam adds foundation models and generative AI next to traditional ML, including agentic workflows. The English MLA-C01 stopped being available after September 28, 2026 (per AWS). Certifications you already earned stay valid for their full term.
How to score yourself
- 9-10 correct: move to full-length, timed practice and review the domains in the exam guide.
- 6-8 correct: reread the task statements for the domains you missed.
- 5 or fewer: this exam assumes hands-on SageMaker AI and Bedrock experience. If you are earlier in your journey, take AWS AI Practitioner (AIF-C01) first. See which AI certification to take first.
How these questions were written
TokIQ Editorial wrote these questions from the domains and task areas in the MLA-C02 exam guide. They are not real exam questions. Exam facts were checked on 2026-11-02, and beta details can change. TokIQ is an independent practice app, not affiliated with or endorsed by Amazon Web Services.
Practice more MLA-C02 questions
TokIQ (on Google Play, coming soon to iPhone) has an MLA-C02 prep mode (beta) with a short daily session, spaced repetition, a readiness score per domain, mock exams and explanations in your own language.
Frequently asked questions
What is MLA-C02?
MLA-C02 is the updated AWS Certified Machine Learning Engineer - Associate exam. It covers traditional ML and foundation models: data preparation, model development, deployment and orchestration, and operating and securing ML and AI solutions (AWS exam guide, checked 2026-11-02).
Is MLA-C02 still a beta exam?
As of 2026-11-02 the AWS certification page says registration is open for the MLA-C02 beta: 85 questions, 170 minutes, 75 USD beta pricing, English only. The general availability date is not announced.
What is the MLA-C02 passing score?
The exam guide reports a scaled score of 100-1,000 with a minimum passing score of 720. The guide notes the pass or fail designation does not apply to the beta version, and the scoring is compensatory, so you pass the exam as a whole, not each domain.
Is MLA-C02 harder than the AI Practitioner exam?
It targets a different level. AIF-C01 is foundational and about choosing the right AI service. MLA-C02 is an associate exam for people with at least 1 year of experience using Amazon SageMaker AI, Amazon Bedrock and other AWS ML services, so expect more hands-on engineering scenarios.
Are these real MLA-C02 exam questions?
No. TokIQ Editorial wrote all 10 from the public exam guide. AWS keeps exam content confidential, so sets sold as real questions are a risk, not a shortcut.
- #MLA-C02
- #AWS Machine Learning Engineer Associate
- #practice questions
- #certification prep