AWS Certified Machine Learning Engineer Associate MLA-C02 Practice Questions with Explana…

154 questions · Cloud Computing · by Study Foxy

Study these questions free

Questions

  1. A training job reads only 5 of 200 columns from a large dataset in S3. Which format reduces I/O the most?

    • Apache Parquet
    • CSV with a header row
    • JSON Lines with one record per line
    • Plain text with tab delimiters

  2. A SageMaker training job should start streaming a huge dataset from S3 right away instead of downloading it all first. Which input mode fits?

    • Fast File mode or Pipe mode to stream the data
    • File mode with a larger EBS volume
    • Batch transform with a manifest file
    • Local mode on a notebook instance

  3. Many training instances need shared, low-latency file access to the same dataset of millions of small images. Which storage fits best?

    • Amazon FSx for Lustre
    • Amazon S3 Glacier Instant Retrieval
    • An EBS volume attached to one instance
    • Amazon DynamoDB with on-demand mode

  4. Clickstream events must feed a model's features in near real time. Which ingestion service fits?

    • Amazon Kinesis Data Streams
    • AWS DataSync on a daily schedule
    • Amazon S3 Glacier Deep Archive
    • AWS Snow Family devices

  5. Customer data from S3, RDS, and DynamoDB must be joined into one training dataset with Spark. Which service fits?

    • AWS Glue
    • Amazon Polly
    • Amazon Lex
    • AWS Budgets

  6. A RAG app needs a scalable vector store that supports hybrid keyword and semantic search. Which service fits?

    • Amazon OpenSearch Service
    • Amazon SQS
    • AWS Step Functions
    • Amazon Data Firehose

  7. A team already runs PostgreSQL on RDS and wants to store embeddings there for similarity search. What should it use?

    • The pgvector extension
    • A DynamoDB TTL attribute
    • An S3 lifecycle rule
    • A Kinesis enhanced fan-out consumer

  8. Features must be served to a real-time endpoint in milliseconds and also kept historically for training. Which service fits?

    • SageMaker Feature Store
    • SageMaker Ground Truth with a private workforce
    • Amazon Athena with partition projection
    • Amazon Kinesis Video Streams

  9. A model will train on live camera footage. Which service ingests and stores it?

    • Amazon Kinesis Video Streams
    • Amazon SQS FIFO queues
    • Amazon Keyspaces
    • AWS Glue DataBrew

  10. A training job fails because its EBS volume fills up while copying data. What is the most direct fix?

    • Increase the volume size or stream the data instead
    • Lower the learning rate of the model
    • Add more epochs to the training job
    • Turn on S3 Versioning for the bucket

  11. SageMaker built-in algorithms can accept RecordIO-protobuf data for faster training in Pipe mode.

    True or false

  12. A feature ranges from 1 to 1,000,000 and is heavily right-skewed. Which transformation reduces the skew?

    • Log transformation
    • One-hot encoding
    • Label encoding
    • Removing the column's header

  13. Features must be rescaled so each has a mean of 0 and standard deviation of 1. What is this called?

    • Standardization
    • Min-max normalization
    • One-hot encoding
    • Binning

  14. A color feature has values red, green, and blue with no order. How should it be encoded for a linear model?

    • One-hot encoding
    • Label encoding as 1, 2, and 3
    • Log transformation
    • Standardization

  15. Ages should be grouped into ranges such as 18-25 and 26-35. Which technique is this?

    • Binning
    • Tokenization
    • Standardization
    • Feature hashing

  16. A data scientist wants a visual tool in SageMaker to explore, transform, and export data for training. Which fits?

    • SageMaker Data Wrangler
    • SageMaker Neo
    • SageMaker Model Monitor
    • SageMaker Debugger

  17. Documents for a RAG app are long, and retrieval returns irrelevant passages. Which change usually helps most?

    • Use a better chunking strategy, such as semantic chunks
    • Raise the model's temperature setting
    • Store the documents as one large chunk
    • Remove all metadata from the documents

  18. Text and images must be turned into vectors for similarity search. What produces these vectors?

    • An embedding model
    • A tokenizer alone
    • A data quality rule
    • A NAT gateway

  19. Training data contains names and Social Security numbers that the model doesn't need. What should happen before training?

    • Mask, redact, or remove the PII fields that aren't needed
    • Encrypt the data and train on it as is
    • Duplicate the records for robustness
    • Move the data to a public bucket

  20. A company has a large corpus of unlabeled domain text and wants an FM to learn its vocabulary. Which dataset preparation fits?

    • Prepare raw text for continued pre-training
    • Prepare prompt-response pairs for instruction tuning
    • Prepare image labels for object detection
    • Prepare time-series data for forecasting

  21. Streaming records need a light transformation before landing in a feature pipeline. Which compute fits for simple per-record changes?

    • AWS Lambda
    • Amazon EMR on a long-running cluster
    • SageMaker HyperPod
    • AWS Snow Family

  22. Breaking text into words or subwords before modeling is called what?

    • Tokenization
    • Binning
    • Normalization
    • Imputation

  23. A feature contains a timestamp, but the model needs day-of-week and hour separately. Which technique is this?

    • Feature splitting
    • Feature hashing
    • One-hot decoding
    • Label smoothing

  24. A team needs thousands of images labeled by humans with built-in quality controls. Which service fits?

    • SageMaker Ground Truth
    • SageMaker Clarify
    • SageMaker Neo
    • Amazon Inspector

  25. Only 1% of transactions in the training data are fraud. Which technique helps the model learn the minority class?

    • Oversample the minority class, such as with SMOTE
    • Remove all fraud examples from the dataset
    • Increase the learning rate tenfold
    • Train only on the majority class

  26. Before model building, a team wants to measure whether one demographic group is underrepresented in the labels. Which tool fits?

    • Clarify pre-training bias metrics
    • SageMaker Debugger built-in rules
    • SageMaker Neo compilation
    • Amazon CloudWatch alarms

  27. A Glue pipeline must stop if a feature has more than 5% missing values. Which feature fits?

    • AWS Glue Data Quality
    • AWS Glue job bookmarks
    • AWS Glue Flex execution
    • AWS Glue crawlers

  28. A numeric feature has a few missing values and outliers. Which imputation is most robust to the outliers?

    • The median
    • The mean
    • The maximum
    • Zero for every row

  29. Data for a time-series forecasting model must be split into train and test sets. How should it be split?

    • Chronologically, testing on later data
    • Randomly shuffled across all dates
    • Alphabetically by record ID
    • By keeping only weekends in training

  30. A fine-tuning dataset of prompt-response pairs may include toxic or mismatched pairs. What should be done before training?

    • Validate and safety-screen the pairs
    • Increase the number of epochs to compensate
    • Remove the system prompt from the model
    • Skip validation to save time

  31. Duplicate records appear in both training and test sets. What problem does this cause?

    • Leakage that inflates test scores
    • Slower inference on the endpoint
    • Higher S3 storage class costs
    • Lower GPU memory usage

  32. Which statement about data augmentation for an image classifier is correct?

    • It creates varied copies, such as flips, crops, and rotations
    • It deletes images that are hard to classify
    • It replaces the labels with random values
    • It converts images into tabular CSV rows

  33. Which techniques help address class imbalance? (Select TWO.)

    • Oversampling the minority class
    • Using class weights in the loss function
    • Removing the minority class entirely
    • Increasing the batch size only

  34. Shuffling data before splitting is appropriate for most non-time-series datasets.

    True or false

  35. Audio files, images, and text transcripts for a multimodal model must be stored cheaply and durably at scale. Which service fits?

    • Amazon S3
    • Amazon ElastiCache for session data
    • Amazon EBS volumes on each training node
    • Amazon DynamoDB items with binary attributes

  36. A training job reports that a DynamoDB export is too slow and throttles the table. What reduces impact on production traffic?

    • DynamoDB export to S3
    • Run full table Scans every hour during peak traffic
    • Lower the table's provisioned read capacity
    • Read the table with strongly consistent reads only

  37. Streaming records must be written to the Feature Store's online store as they arrive. What does the pipeline call?

    • The PutRecord API
    • The Glue StartCrawler API
    • The S3 RestoreObject API
    • The KMS GenerateDataKey API

  38. A team must label-encode an ordinal feature with values low, medium, and high. Which encoding preserves the order?

    • Ordered integers: 0, 1, 2
    • One-hot encoding with no ordering
    • Hashing the strings into random buckets
    • Dropping the feature entirely

  39. Large-scale feature engineering on terabytes of data needs Spark. Which service fits?

    • Amazon EMR
    • Amazon Lex
    • AWS Budgets
    • Amazon Polly

  40. Chunks retrieved for a RAG app need filters such as department and document date. What should be extracted during preparation?

    • Per-chunk metadata
    • A larger embedding dimension for every chunk
    • A higher temperature for the generator
    • A single chunk for the whole corpus

  41. A dataset must be prepared to distill a large FM's behavior into a smaller model. What does the dataset typically contain?

    • Prompts with the teacher's responses
    • Only random noise inputs with no responses
    • Raw images with no labels or prompts
    • Duplicate rows from the test set

  42. Analysts want no-code profiling to find missing values and outliers in a dataset. Which service fits?

    • AWS Glue DataBrew
    • AWS DataSync
    • Amazon Route 53
    • AWS CloudHSM

  43. Removing outliers is always the correct way to handle them.

    True or false

  44. Several different models with different frameworks must be served from one endpoint, each invoked directly. Which option fits?

    • A multi-container endpoint
    • A multi-model endpoint that shares one container
    • Batch transform with one container per job
    • Serverless inference with one framework only

  45. Which statement about Bedrock cross-Region inference is accurate?

    • It routes requests to other Regions to handle traffic bursts
    • It copies training data to every AWS Region
    • It disables Guardrails for faster responses
    • It requires a separate account in each Region

  46. A retrieval pipeline must answer within 300 ms to meet an SLA. Which change most directly reduces retrieval latency?

    • Fewer, smaller chunks from a tuned index
    • Retrieve every chunk in the index for each query
    • Use the largest embedding model available
    • Add more documents to each chunk

  47. Automated retraining should deploy the new model without checking its evaluation metrics.

    True or false

  48. Model Monitor must compare which features drive predictions now versus at training time. Which monitor type fits?

    • Feature attribution drift monitoring
    • Data quality monitoring of null counts
    • EBS volume monitoring
    • CloudTrail management events

  49. Many low-traffic models each run on their own endpoint, and most instances sit idle. What reduces cost?

    • A shared multi-model endpoint
    • Give each model a larger instance
    • Add a second endpoint per model
    • Turn off auto scaling on every endpoint

  50. SageMaker execution roles should usually include AdministratorAccess so jobs never fail.

    True or false

Showing 50 of 154. Open the quiz to study them all.