A team is training a model on a dataset with 100,000 samples. They set the batch size to 1,000 and train for 10 epochs. The training job takes 5 hours. They want to reduce the training time while keeping the same number of epochs. What should they do?

Increase the batch size to 2,000, which reduces the number of steps per epoch.

Keep the batch size the same but reduce the learning rate.

Decrease the batch size to 500, which increases the number of steps per epoch.

Increase the number of epochs to 20.

A company wants to build a sentiment analysis model for customer reviews. They have a small labeled dataset of 5,000 reviews. They want to use a pre-trained transformer model and fine-tune it on their data, then deploy it for real-time inference. Which approach should they take?

Use Amazon Bedrock to fine-tune a foundation model, but deploy the model using a SageMaker endpoint.

Use SageMaker JumpStart to fine-tune a pre-trained Hugging Face BERT model on the dataset, then register the model in SageMaker Model Registry and deploy it to a SageMaker endpoint.

Train a custom neural network from scratch using SageMaker script mode with PyTorch, without any pre-trained weights.

Use SageMaker built-in BlazingText algorithm to train a text classification model and deploy it directly without versioning.

An ML team is training a large transformer model on SageMaker. The training job is expected to take 20 hours on a single GPU instance. The team has a budget that allows them to use up to 4 GPU instances. They also need to ensure the model converges well and does not overfit. They have already implemented early stopping. Which strategy will reduce training time the most while maintaining model quality?

Use SageMaker's data parallelism library to distribute the training across 4 GPU instances, and increase the batch size proportionally.

Increase the learning rate and reduce the number of epochs to speed up training.

Use a single larger GPU instance with more memory and keep the training script unchanged.

Use SageMaker's model parallelism library to split the model across 4 GPU instances without changing the batch size.

What is the first step to bring a model trained outside SageMaker into SageMaker for deployment?

Create a SageMaker notebook instance to re-train the model.

Convert the model to a SageMaker built-in algorithm.

Save the model artifacts (e.g., weights and architecture) to an Amazon S3 bucket.

Register the model in AWS Marketplace.

A deep learning model for speech recognition is overfitting. The team has already tried increasing the dropout rate, but the validation accuracy is still not improving. They want to apply another regularization technique that penalizes large weights. Which technique should they add?

Add more layers to the network.

Add L2 weight decay to the optimizer.

AWSCertified Machine Learning Engineer - Associate

Domain 2Objective 2

Task 2.2: Train and Refine Models MLA-C01 Practice Questions (Page 2)

Part of the Content Domain 2: ML Model Development domain, which accounts for 26% of the MLA-C01 exam. AWS does not publish an official question count, but from its 130-minute exam (~50–85 total, ~13–22 in this domain), expect 4–7 from this objective — we provide 43 practice questions to prepare you well beyond it. (estimate)

43questions here

9free pages

16concepts

26%of the exam

16 Concepts in This Objective

Training process elements
Reducing training time
Factors influencing model size
Improving model performance
Regularization techniques
Hyperparameter tuning techniques
Model hyperparameters and effects
Integrating external models into SageMaker
SageMaker built-in algorithms and libraries
SageMaker script mode
Fine-tuning pre-trained models
SageMaker automatic model tuning
Preventing overfitting and underfitting
Ensembling and model combination
Reducing model size
Model version management

Questions 6–10

6application · medium
A team is training a model on a dataset with 100,000 samples. They set the batch size to 1,000 and train for 10 epochs. The training job takes 5 hours. They want to reduce the training time while keeping the same number of epochs. What should they do?
Select an answer first
7application · medium
A company wants to build a sentiment analysis model for customer reviews. They have a small labeled dataset of 5,000 reviews. They want to use a pre-trained transformer model and fine-tune it on their data, then deploy it for real-time inference. Which approach should they take?
Select an answer first
8expert · hard
An ML team is training a large transformer model on SageMaker. The training job is expected to take 20 hours on a single GPU instance. The team has a budget that allows them to use up to 4 GPU instances. They also need to ensure the model converges well and does not overfit. They have already implemented early stopping. Which strategy will reduce training time the most while maintaining model quality?
Select an answer first
9foundation · easy
What is the first step to bring a model trained outside SageMaker into SageMaker for deployment?
Select an answer first
10application · medium
A deep learning model for speech recognition is overfitting. The team has already tried increasing the dropout rate, but the validation accuracy is still not improving. They want to apply another regularization technique that penalizes large weights. Which technique should they add?
Select an answer first

Finished these 5 questions?

Review the revealed explanations, or continue through the curriculum.

Next 5 questions →Back to MLA-C01

Free Basic Practice is a study aid with revealable answers — not a scored exam. Examers.io is independent and not affiliated with or endorsed by AWS. “MLA-C01” is a trademark of its owner, used for identification only.