Free AWS MLA-C01 Actual Exam Questions

Probably C. Increasing target_precision is the only option directly tied to reducing false positives since precision measures how many predicted positives are actually correct. A and B don’t really impact false positive rates directly, just model complexity or training time. D makes no sense because switching to a regressor won’t help with classification errors like false positives. The question is about minimizing false positives, and focusing on precision is the most straightforward way.

C imo, since increasing target_precision directly reduces false positives by emphasizing precision.

Guessing A, LightGBM handles imbalanced data and complex feature interactions nicely.

Good point about clustering and topic modeling being off for fraud detection, so A it is.

D. Since predicting prices is a regression task, classification metrics like accuracy or F1 don't apply. MAE gives a clear average error magnitude, making it the most suitable here.

D. Since this is about predicting apartment prices, which is a continuous value, you want a regression metric. Accuracy, AUC, and F1 are for classification tasks, so they’re not relevant here. MAE is straightforward—it tells you the average absolute difference between predicted and actual prices, which makes the model’s performance easier to interpret in real-world terms.

Option C seems simpler here since it avoids extra setup like Spark or local downloads, and you can quickly write a small script to drop columns using SageMaker processing.

C, since it avoids setting up Spark or local downloads and is simpler than EMR.

A, because Secrets Manager handles rotation natively, unlike KMS options here.

A imo, because Secrets Manager is built for secret rotation with built-in Lambda support, making it easier to automate token updates safely every 3 months. Parameter Store isn’t as specialized for rotation.

Parquet files are definitely optimized for complex data and fast processing, so D still makes sense here. CSV and JSON just don’t handle complexity or speed as well. D

A vs D? Parquet usually beats CSV for complex data, but codec support might mess things up.

A/B? A sounds solid since SageMaker Clarify is built for bias detection and can run on demand, which matches the requirement well. But B might still work if you want more control by pulling the model-monitor-analyzer image and running custom checks via Lambda. The question asks for on-demand bias drift monitoring, so both have merits, but Clarify feels more straightforward for this use case.

Maybe B since pulling the built-in analyzer image lets you customize bias checks flexibly.

Adeel M.: A imo—IAM policies on Glue catalog directly control Athena queries without extra S3 policy complexity.

It’s C—Lake Formation handles both S3 and Athena access in one place efficiently.

Feature splitting doesn’t seem necessary here since no combined features are mentioned. One-hot encoding fits well for categorical variables, and standardization is good for numeric features with a normal distribution.

I think feature splitting can be ruled out since there’s no mention of combined features like “city-state” or something that needs breaking apart. One-hot encoding is a no-brainer for categorical data like neighborhood or style of house, because numeric models don’t handle categories well. For numeric features, if they're roughly normal, standardization works best, but if the distribution is skewed (like price or area), applying a log transform helps normalize it before feeding into the model. So, the three techniques should be one-hot encoding, log transform, and standardization.

Embedding, RAG, and Temperature fit best since Token is basic input unit.

Embedding, RAG, and Temperature cover more specific generative AI roles than Token.

Starting with creating the feature group (B) is essential since it defines the schema. Then ingesting records (C) populates the store, and finally accessing the store (A) lets you build datasets for training.

Starting with creating the feature group sets up the schema and storage. Next, ingesting records fills it with data. Finally, accessing the store lets you build training datasets from those features.

I’m thinking AWS Glue crawlers (D) should come first to detect any schema info despite no headers, since Athena usually relies on some metadata or headers to infer columns. After that, DataBrew (E) seems like the best move to clean and engineer features visually without coding. Finally, storing the cleaned data back in S3 (B) wraps it up. The batch transform job feels more like model inference than prepping raw unlabeled data. This order makes sense for a smooth pipeline from raw to ready-for-ML data.

I’d go with using AWS Glue crawlers (D) first, even without headers, since they can still detect schema patterns. Then use DataBrew (E) to clean and engineer features because it’s designed for that kind of task on raw data. Finally, store the cleaned dataset back in S3 (B) so it’s ready for model training. Batch transform feels more like a later step after you have clean, labeled data, not for initial prep. Athena is more for querying directly, less about preparing or cleaning the data itself.

The pipeline should start with the S3 event notification to trigger it, then SageMaker retrains the model using the new data, and finally, it deploys the updated model to the SageMaker endpoint for real-time use.

S3 event notification is definitely the trigger, so that removes the lifecycle rule option. After retraining, deploying to Model Registry makes sense before endpoint deployment, but since only three steps are allowed, I’d pick S3 event, retrain, then deploy to Model Registry.

C imo, model groups in SageMaker Model Registry are designed to organize models and versions efficiently, which cuts down on manual tracking and operational work compared to just tagging versions.

Maybe D makes more sense here. The question emphasizes versioning within a central registry, and using unique tags for each version is straightforward without adding complexity from managing model groups. Model groups are good for big catalogs, but if the setup is simpler, tagging alone might keep things leaner. Plus, both C and D use SageMaker Model Registry, so the core requirement is met either way, but D could minimize overhead by avoiding extra organization layers.

Makes sense to skip polling since it's inefficient. C should be the easiest — EventBridge triggers right on the upload event, so no extra Lambda or Airflow setup needed. Definitely less to maintain overall. C

Probably C here. EventBridge can listen for S3 object-created events and trigger the SageMaker pipeline directly, so you avoid any extra Lambda functions or complicated orchestration setups. That’s less work to manage over time. A and D sound like overkill since Lifecycle rules aren’t meant to trigger training and MWAA is more for complex workflows, not simple event triggers. B works but has more overhead with the Lambda polling or scanning, which isn’t necessary given EventBridge’s native integration with S3 events.