CertifyDeckCargando contenido…
MLA-C02 · Sample questions
10 real questions from the CertifyDeck MLA-C02 bank, spread across the domains of the official AWS Certified Machine Learning Engineer - Associate skills outline, with the correct answer and a full explanation of why the other options are wrong. They are original questions written from the public syllabus, not exam dumps, and they are the same style you will find in the practice tests and mock exams.
Correct answer: B. Text Classification - TensorFlow
Text Classification - TensorFlow is a supervised built-in algorithm that supports transfer learning with available pretrained models, which suits assigning documents to predefined categories with a small labeled dataset. LDA and NTM are unsupervised topic modeling algorithms that discover topics not known in advance, so they cannot map documents to the firm's 12 fixed practice areas. Sequence-to-Sequence is meant for tasks such as machine translation and summarization, not classification.
Correct answer: B. Model parallelism, so that the model is partitioned across the GPUs in the cluster.
Model parallelism partitions the model across multiple GPUs, so it is the strategy for models that are too large to fit in the memory of a single device. Data parallelism still requires that at least the full model and one record fit on each GPU, because every GPU holds a complete replica. Warm pools only shorten provisioning time, and sharding the input data does not reduce the memory that the model itself needs.
Correct answer: D. Enable prompt caching and place the contract before the questions as a static prefix.
Prompt caching reduces response latency and input token costs for workloads with long contexts that are reused across queries, such as a document that users question repeatedly; static content should come first so later requests match the cached prefix. The Priority tier gives faster responses but at a price premium, so it raises cost. Batch inference is asynchronous and does not support prompt caching, and the Flex tier trades longer processing times for its discount.
Correct answer: D. A split by key on the customer ID column so that each customer appears in only one set.
The split by key transform prevents rows with the same key from occurring in more than one split, so no customer's history leaks from training into testing and the evaluation reflects performance on unseen customers. A randomized split takes random non-overlapping samples of rows, which still scatters a customer's rows across both sets. A stratified split balances label proportions but does not group rows by customer, and an ordered split follows the row order, which does not guarantee that a customer's rows stay together.
Correct answer: B. Run an Amazon Comprehend asynchronous PII analysis job on the S3 prefix with redaction of the PII entities.
Amazon Comprehend detects PII in English or Spanish text, and redacting PII entities requires an asynchronous batch job that processes documents in Amazon S3. Real-time analysis can only locate PII entities, so it would not return redacted text. Macie discovers sensitive data and generates findings without modifying objects, and a Glue crawler only infers schemas.
Correct answer: D. Write the prompts to JSONL files in Amazon S3 and submit a batch inference job that writes results to S3.
Amazon Bedrock batch inference processes a large number of prompts asynchronously from input files in Amazon S3 with a single job request and writes the responses back to S3, which fits a large offline labeling task. A Lambda loop of individual InvokeModel calls means building retries, throttling handling and orchestration yourself. Batch inference is not supported for provisioned models, and hosting a model on a real-time endpoint adds infrastructure that an offline job does not need.
Correct answer: D. Submit a fine-tuning job whose base model is the previous custom model, and give the output a new custom model name.
A model customization job accepts either a foundation model or a previously customized (fine-tuned or distilled) model as its base, and it creates a new custom model with the name you give it, so earlier versions remain available for rollback. Fine-tuning the base foundation model again does not start from the previous version, and replacing the earlier model removes the rollback target. Custom Model Import brings in models customized outside Amazon Bedrock, such as in SageMaker AI, and distillation transfers knowledge from a teacher model to a smaller student model instead of continuing the previous version's training.
Correct answer: A. Retrieve, which returns the most relevant source chunks as an array without generating a response from them
Retrieve returns the most relevant source chunks as an array, which lets the application decouple the RAG steps and apply its own processing and orchestration before calling a model. RetrieveAndGenerate and its streaming variant run retrieval and generation together, so the in-house rules could only be requested in a prompt rather than enforced in code before generation. GenerateQuery converts natural language into queries for structured data stores and does not return document chunks.
Correct answer: B. The extra time is added by SageMaker AI outside the container, for example by larger payloads or authentication.
OverheadLatency measures the time SageMaker AI adds between receiving the request and returning the response, minus ModelLatency, and it varies with payload sizes, request frequency, and authentication or authorization. Because ModelLatency, which covers inference inside the container, is unchanged, the container is not the cause, so optimizing the model or enlarging instances would not help. More 5xx errors would show in Invocation5XXErrors, and model download and load times are reported by separate multi-model endpoint metrics.
Correct answer: B. Create an EventBridge rule for SageMaker pipeline step status change events with a Failed status that targets an SNS topic.
SageMaker AI sends pipeline step status change events to EventBridge in near real time, and when the step status is Failed the event includes the step name and a failureReason field, so a rule that targets SNS delivers both details without custom code. An alarm on ExecutionFailed signals that an execution failed but does not carry the failed step or the reason. A polling Lambda function adds code to maintain and is not near real time, and anomaly detection on duration does not detect failures.
Answer each question before opening the explanation, and write down why you discarded the other options. On MLA-C02, most mistakes come from choosing an option that is technically possible but does not meet every requirement in the scenario.
These 10 questions are a taste of the format, not a diagnostic. To know which domains you fail, practice by domain for free and take a timed mock exam that mirrors the official weights and duration.
Start for free, no sign-up or card. Every question includes an explanation like the ones above, and the mock exam reproduces the official MLA-C02 format.