Free MLA-C02 practice test: 20 questions
Twenty questions across the four MLA-C02 domains, which carry almost equal weight on the real exam. These are engineering questions: the right answer is usually the one that survives production, not the one that is quickest to build.
Data Preparation for ML and AI
Question 1. A training dataset has a feature that is missing for 60% of rows and correlates weakly with the target. What is usually the best action?
- A. Impute the missing values with the column mean
- B. Impute with zero and keep the feature
- C. Drop the feature
- D. Drop every row with a missing value
Show answer
Answer: C
A feature missing for most rows and weakly related to the target adds noise and imputation risk for little gain, so dropping it is usually right. Imputing 60% of values invents most of the column.
Question 2. A fraud dataset contains 0.3% positive cases. Training produces a model with 99.7% accuracy that never predicts fraud. What should change first?
- A. Collect more data of the same distribution
- B. Use imbalance-aware metrics and apply resampling or class weighting
- C. Increase the model size
- D. Reduce the number of features
Show answer
Answer: B
With severe class imbalance, accuracy is meaningless and must be replaced by metrics like precision, recall or AUC-PR, alongside resampling or class weighting. Collecting more of the same imbalanced data does not fix the metric problem.
Question 3. A feature computed from the target variable's future value leaks into training. What is the symptom at deployment?
- A. Training fails to converge
- B. Training is unusually slow
- C. Offline and production performance match closely
- D. Excellent offline metrics and poor production performance
Show answer
Answer: D
Target leakage produces excellent offline scores that collapse in production, because the leaked information is unavailable at inference time.
Question 4. Training and serving compute the same feature with slightly different code, and predictions differ from expectations. What is this called?
- A. Training-serving skew
- B. Concept drift
- C. Overfitting
- D. Class imbalance
Show answer
Answer: A
Training-serving skew arises when feature computation differs between the two paths. A feature store with shared transformations is the standard remedy.
Question 5. Personal data must be removed from a training set while keeping the records usable for modelling. Which approach is appropriate?
- A. Delete every row containing personal data
- B. Encrypt the dataset at rest and train on the decrypted copy
- C. Pseudonymise or tokenise the identifying fields
- D. Rename the columns
Show answer
Answer: C
Pseudonymisation or tokenisation of identifiers keeps row-level structure without exposing identities. Deleting all rows destroys the dataset, and encryption alone still yields identities once decrypted for training.
ML Model and Foundation Model Development
Question 6. Cross-validation scores vary widely between folds on a small dataset. What does this most likely indicate?
- A. The model is production-ready
- B. The performance estimate is unstable and needs repeated or stratified validation
- C. The learning rate is too low
- D. The dataset is too large
Show answer
Answer: B
High variance across folds on limited data indicates the estimate itself is unstable, calling for repeated or stratified cross-validation and caution about the result. It is not evidence of a good model.
Question 7. For a medical screening model, missing a positive case is far worse than a false alarm. Which metric should be prioritised?
- A. Recall
- B. Precision
- C. Accuracy
- D. Specificity only
Show answer
Answer: A
Recall measures the proportion of true positives captured, which is exactly what must not be missed. Precision penalises false alarms, which are the acceptable cost here.
Question 8. Hyperparameter tuning must explore a large space under a fixed compute budget. Which strategy is generally most efficient?
- A. Exhaustive grid search over every combination
- B. Training one model with default settings
- C. Manual tuning by intuition
- D. Bayesian or adaptive search with early stopping of poor trials
Show answer
Answer: D
Bayesian or adaptive search concentrates trials in promising regions and prunes poor ones, using the budget better than exhaustive or uninformed search.
Question 9. A foundation model must answer from proprietary documents. Fine-tuning is rejected on cost grounds. What fits the constraint?
- A. Retrieval-augmented generation over an index of the documents
- B. Continued pre-training on the documents
- C. Training a model from scratch
- D. Full fine-tuning on the documents
Show answer
Answer: A
Retrieval-augmented generation avoids any training cost by supplying documents at inference. Continued pre-training is the most expensive option of all.
Question 10. Two candidate models perform within a percentage point of each other, but one is far simpler and faster. Which should usually ship?
- A. The more complex model, because it scores higher
- B. Neither, until a third is trained
- C. The simpler, faster model
- D. Both, served in parallel permanently
Show answer
Answer: C
When performance is effectively equal, the simpler and cheaper model is easier to operate, explain and maintain. Marginal metric gains rarely justify operational complexity.
Question 11. Training a large model repeatedly exhausts the budget. Which change reduces cost with least risk to quality?
- A. Halve the training dataset
- B. Use interruptible capacity with regular checkpointing
- C. Stop training after one epoch regardless of convergence
- D. Remove validation entirely
Show answer
Answer: B
Spot or interruptible capacity with checkpointing cuts cost substantially for fault-tolerant training. Reducing the dataset or stopping early both risk quality directly.
Deployment and Orchestration of ML and AI Workflows
Question 12. A new model version must be exposed to 5% of traffic before full rollout, with instant rollback available. Which pattern applies?
- A. Canary deployment with weighted traffic routing
- B. Replace the endpoint in place
- C. Deploy to all traffic and monitor closely
- D. Deploy only in the development account
Show answer
Answer: A
Canary deployment routes a small traffic share to the new version and allows immediate rollback. Replacing the endpoint outright removes the safety net.
Question 13. Predictions are needed for 40 million records once a night, with no latency requirement. Which serving mode is appropriate?
- A. A real-time endpoint running continuously
- B. A serverless endpoint per record
- C. A scheduled batch inference job
- D. An interactive notebook
Show answer
Answer: C
Batch transform processes large volumes efficiently and costs nothing between runs. A real-time endpoint would run continuously for no benefit.
Question 14. A pipeline must retrain automatically whenever new labelled data lands, with full lineage recorded. What should orchestrate this?
- A. A daily reminder to a data scientist
- B. A notebook run manually when someone remembers
- C. A cron job that copies files
- D. An event-triggered ML pipeline with artefact and lineage tracking
Show answer
Answer: D
An event-triggered pipeline with tracked artefacts and lineage gives both automation and auditability. Manual notebook runs record nothing reliably.
Question 15. An endpoint sees ten times the traffic at month end. Costs must stay low the rest of the month. What should you configure?
- A. Autoscaling with appropriate minimum and maximum capacity
- B. Permanently provision for peak load
- C. Manually scale up when someone notices
- D. Throttle all requests to a fixed rate
Show answer
Answer: A
Autoscaling adds capacity under load and removes it afterwards, matching cost to demand. Permanently provisioning peak capacity wastes money for most of the month.
Question 16. A model artefact in production cannot be traced to the code and data that produced it. What practice was missing?
- A. Endpoint request logging
- B. Model registry entries versioned with code, data and parameters
- C. Cost allocation tags
- D. A larger instance type
Show answer
Answer: B
Versioning models alongside the code, data and parameters that produced them gives reproducibility. Endpoint logging records use, not provenance.
Operating, Monitoring, and Securing ML and AI Solutions
Question 17. A model's accuracy degrades over six months although the code has not changed. What should monitoring detect?
- A. Increased request latency
- B. Higher HTTP error rates
- C. Data and concept drift against the training baseline
- D. Rising storage costs
Show answer
Answer: C
Data and concept drift describe the input distribution or the target relationship shifting away from training conditions. Latency and error rates would not reveal this.
Question 18. Ground-truth labels arrive two weeks after predictions. How should model quality be monitored in the meantime?
- A. Monitor input and prediction distributions as proxies, then compute true metrics when labels arrive
- B. Wait two weeks and do nothing in the interim
- C. Assume performance is unchanged
- D. Retrain daily regardless
Show answer
Answer: A
Proxy signals such as prediction distribution and input drift give early warning before labels arrive, with true performance computed later. Waiting passively leaves two weeks blind.
Question 19. A training job needs read access to one S3 prefix and nothing else. What should you configure?
- A. An access key belonging to an administrator
- B. Full read access to all buckets in the account
- C. Public read access on the bucket
- D. An execution role granting read access to that prefix only
Show answer
Answer: D
A role scoped to exactly that prefix follows least privilege. Broad account access or long-lived user keys grant far more than the job needs.
Question 20. An inference endpoint must not be reachable from the public internet, but an internal service must call it. What applies?
- A. A public endpoint with an IP allow-list
- B. A VPC-private endpoint reachable only from inside the network
- C. Basic authentication on a public endpoint
- D. Disabling the endpoint entirely
Show answer
Answer: B
A private endpoint inside the VPC keeps traffic off the internet while remaining reachable internally. IP allow-lists still expose a public endpoint.
How did you do?
Sixteen or more correct suggests you are close to ready. Below fourteen, the domain guides in this section are the fastest route back — MLA-C02 leans heavily on operational judgement rather than algorithm theory.