Independent: not affiliated with AWS. Free to try; the study bundle adds a study guide and two practice exams.
5 original questions, none of them in the practice exams.
Pick an answer. Show answer explains every option and links its AWS source.
5 questions / Explanations link AWS sources / Last updated Oct 26, 2026
Question 1
Multiple choice
Domain 1 / Task 1.1 Collect and store data
A hospital research group keeps 80 TB of de-identified imaging studies on an on-premises NFS array, and new studies are added every day. An Amazon SageMaker AI training pipeline needs each day's new files in an Amazon S3 bucket by the next morning. The transfer must be encrypted in transit, must detect any file that does not arrive intact, and must run without the team writing or operating transfer scripts. Which service should the engineer use?
Show answerHide answer
Answer: A
Why
DataSync transfers file and object data between on-premises storage (NFS, SMB, HDFS and object storage) and AWS storage services including Amazon S3, Amazon EFS and Amazon FSx. It includes encryption and data integrity validation so data arrives intact, uses a purpose-built protocol with a parallel, multi-threaded architecture, and AWS names recurring data-processing workflows for machine learning among its use cases. A task schedule, defined with a cron or rate expression at a minimum interval of one hour, runs the transfer each night without a script.
Why not the others
BAWS Direct Connect is a network connection between a site and AWS; Amazon S3 does not read files from an NFS export over it.
CA cron job and CLI script are exactly the transfer code the team would have to write and operate, which the requirement rules out.
DAmazon Data Firehose delivers streaming records from producers to destinations; it does not read files from an on-premises file system.
A furniture retailer is adding visual search to a catalog of 300,000 product photos. A shopper must be able to type a description such as 'walnut mid-century sideboard' or upload a photo of a piece seen elsewhere, and either kind of query must search the same index of product-image vectors. Which approach should the engineer use to produce the vectors?
Show answerHide answer
Answer: B
Why
Amazon Titan Multimodal Embeddings G1 translates text, images, or a combination of text and images into embeddings in a shared semantic space, and AWS names searching for images by text or by image similarity as its use case. Embedding every product photo and every query with this one model puts text queries and photo queries into the same space as the index. The model takes up to 256 input text tokens and images up to 25 MB at up to 2048 x 2048 pixels, and returns vectors of 1,024 (default), 384 or 256 dimensions.
Why not the others
AThe pre-trained model can be used as-is; fine-tuning on image-text pairs is an optional customization, not a prerequisite for search.
CAmazon Titan Text Embeddings V2 takes a body of text as input and returns a vector; it cannot embed a shopper's uploaded photo.
DVectors produced by two different models do not share a semantic space, so a cosine similarity between them does not measure how alike an image and a description are.
Domain 2 / Task 2.2 Train, fine-tune and customize
In June, a team ran a 150-job Amazon SageMaker AI automatic model tuning job for a demand-forecasting model, with validation:rmse as the objective metric. It now wants a new tuning job that starts from the June results rather than from scratch. Since June, planners have started judging forecasts by absolute error, the training set has grown by 40 percent, seven older tuning jobs for related models have finished, the best June training job used the lowest learning rate in the range searched, and only a few batch sizes run well on the training instances. The team has drafted one change for each of these developments. Which two of the drafted changes below can a warm start tuning job of type TRANSFER_LEARNING accept? Choose two.
Show answerHide answer
Answers: A and B
Why
A warm start tuning job uses one or more previous tuning jobs as a starting point, and their results inform which hyperparameter combinations the new job searches. The TRANSFER_LEARNING type allows the input data, hyperparameter ranges, maximum number of training jobs and maximum number of concurrent training jobs to differ from the parent jobs, and even a different version of the training image, so training on the enlarged dataset and searching below the old lower bound for the learning rate are both permitted. AWS lists tuning a model on newly received data and changing hyperparameter ranges among the reasons to use warm start, with the caveat that dataset changes that significantly affect the objective metric might reduce how useful warm start is.
Why not the others
CThe objective metric in the new tuning job must be the same as the objective metric used in the parent jobs, so validation:rmse cannot become validation:mae.
DThe type of each hyperparameter (continuous, integer or categorical) must not change between the parent jobs and the new tuning job, so an integer range cannot become a categorical list.
EA tuning job can have a maximum of 5 parent jobs, and the June job plus seven older jobs makes 8.
Domain 3 / Task 3.3 Automate orchestration and CI/CD
An ML team maintains one Amazon SageMaker Pipelines definition for a churn model. In development it must process a 2 GB sample from one S3 prefix on ml.m5.xlarge instances and register the model as PendingManualApproval; in production it must process the full 400 GB dataset from another prefix on ml.m5.4xlarge instances and register the model as Approved. The team does not want a second pipeline to maintain, or an edit to the definition between runs. What should the engineer do?
Show answerHide answer
Answer: D
Why
Pipeline parameters introduce variables into a pipeline definition. Each parameter has a type (ParameterString, ParameterInteger, ParameterFloat or ParameterBoolean) and a default value, can be referenced throughout the definition, and has its default overridden by passing parameter values when an execution starts, for example pipeline.start(parameters=dict(ModelApprovalStatus='Approved')). Here the development values are the defaults and the production execution passes its own prefix, instance type and approval status, so one definition serves both environments; every parameter used in a step must be defined in the pipeline's parameters list.
Why not the others
AStep caching reuses the outputs of an earlier successful step run whose inputs match; it does not supply a different prefix, instance type or approval status to an execution.
BA second pipeline is a second definition that has to be kept in sync with the first by hand, which the team has said it does not want.
CA step's script runs on an instance whose type the step definition has already fixed, so a value the script reads at run time cannot choose that instance type.
Domain 4 / Task 4.3 Secure workloads and endpoints
An insurer runs a claims assistant that calls Amazon Bedrock through the Converse API. Its compliance team requires the full prompt and model response of every invocation to be kept in the insurer's own AWS account, where auditors can query them, and the application team must not add logging code to the assistant. Which approach meets the requirement?
Show answerHide answer
Answer: A
Why
Model invocation logging is disabled by default. Once it is enabled for the account and Region, through the console or the PutModelInvocationLoggingConfiguration API, Amazon Bedrock collects the full request data, response data and metadata for Converse, ConverseStream, InvokeModel and InvokeModelWithResponseStream calls, so the application needs no logging code, and publishes them to CloudWatch Logs, Amazon S3 or both; only destinations in the same account and Region are supported. With an S3 destination, each log record carries the caller's identity ARN, the model ID, token counts and request and response bodies up to 100 KB, bodies larger than that are stored as separate objects in the bucket, and the data can be queried with Amazon Athena.
Why not the others
BA trail delivers the events to S3, but the event for a Converse or InvokeModel call records who made the call, when, from which IP address and which model ID was requested; it does not contain the prompt or the model's response.
CA guardrail evaluates content while the call runs and can block or mask it; it is a filter, not a record of prompts and responses.
DAWS Config records the configuration of resources and changes to it, not the requests sent to a model.
The study bundle has two practice exams with 130 original questions, an answer key and an explanation for every answer, plus the 62-page study guide. Practice exam explanations are shorter than the ones on this page: a short paragraph per question, with no source links.
Study bundle
$19
130 practice questions / Study guide / Answer keys
Every bundle includes 6 months of access: download the PDFs, and any new editions, until 6 months after your purchase date. Your receipt has a license key. Enter it at overqual.com/download for the three PDFs, stamped with your email and order number. They are yours to keep.