MLS-C01 Modeling Practice Test 25 — 15 Questions

Question 1

A company is using Amazon SageMaker to deploy a model for real-time inference. The model takes 200 ms to respond, but the requirement is 100 ms. Which THREE actions could reduce latency? (Choose THREE.)

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

Use a larger instance with more compute capacity. Option A is correct because moving to a larger instance with more compute capacity (more vCPUs, higher memory bandwidth, or GPU acceleration) gives the model more resources to execute its forward pass, directly reducing per-inference latency. Option B is correct because pruning removes unnecessary weights and neurons, shrinking the model's computational graph so fewer operations are executed per request, which lowers response time. Option D is correct because SageMaker Neo compiles the trained model into an optimized executable for the specific target instance's hardware (e.g., using instruction sets or accelerators), eliminating framework overhead and speeding up inference. Option C is not appropriate because switching to a CPU-based instance typically increases latency for compute-heavy models compared to GPU or accelerated instances. Option E is not appropriate because increasing the batch size improves throughput by amortizing cost across many requests, but it generally increases the latency of individual responses rather than reducing it.

Answer

Switch to a CPU-based instance

Answer

Increase the batch size for inference

Question 2

A data scientist is performing feature selection for a linear regression model. Which TWO methods are appropriate? (Choose TWO.)

Accepted Answer

Lasso (L1) regularization. Lasso (L1) regularization (A) is appropriate because its L1 penalty drives some coefficients exactly to zero, effectively performing feature selection by shrinking irrelevant predictors out of the linear regression model. Forward selection (D) is also appropriate because it is a classic wrapper-based feature selection method that iteratively adds the predictor that most improves the model until a stopping criterion is met. Ridge (L2) regularization (B) is not a selection method since it only shrinks coefficients toward zero without eliminating any features. t-SNE (C) is a nonlinear dimensionality reduction and visualization technique, not a feature selection method for linear regression. PCA (E) performs feature extraction by creating new uncorrelated components rather than selecting the original features, so it does not meet the feature selection requirement.

Answer

Ridge (L2) regularization

Answer

t-distributed stochastic neighbor embedding (t-SNE)

Answer

Principal component analysis (PCA)

Question 3

Refer to the exhibit. An IAM policy is attached to a SageMaker notebook instance. The data scientist runs a training job that reads from s3://my-bucket/training-data/ and writes to s3://my-bucket/output/. The training job fails with an access denied error. What is the most likely cause?

Accepted Answer

The policy does not allow s3:PutObject on the output location. The training job reads from s3://my-bucket/training-data/ and writes to s3://my-bucket/output/. If the IAM policy grants s3:GetObject on the training prefix but omits s3:PutObject on the output prefix, the job will start successfully (read access works) but fail when it attempts to write model artifacts or output data. That matches the 'access denied' error occurring mid-job rather than at submission time.

Answer

The policy does not allow sagemaker:CreateTrainingJob

Answer

The policy is missing the sagemaker:InvokeEndpoint action

Answer

The policy does not allow s3:GetObject on the training data

Question 4

Refer to the exhibit. A custom training job using Pipe input mode fails. The logs indicate the algorithm cannot read the data. What is the most likely issue?

Accepted Answer

The algorithm expects File mode but Pipe mode is specified. The most likely issue is that the algorithm expects File mode but Pipe mode is specified. In SageMaker, Pipe mode streams data directly from S3 to the training container, but some algorithms are not designed to read from a pipe and require files to be present on the local file system. This mismatch causes the algorithm to fail to read the data.

Answer

The instance type is too small for the data

Answer

The training data is compressed

Answer

The training image is not accessible

Question 5

Refer to the exhibit. A data scientist wants to update the endpoint to use a new model image. The scientist updates the endpoint configuration with the new image and calls UpdateEndpoint. After the update, the endpoint status is 'Updating' but remains in that state for a long time. What is the most likely cause?

Accepted Answer

The new model image is failing health checks. When an endpoint update stalls in 'Updating' status, the most likely cause is that the new model image is failing health checks, so SageMaker cannot bring the new instances into service. The endpoint remains in 'Updating' while it retries or waits for healthy instances. This is a common failure mode when the model container doesn't respond correctly to the /ping health check endpoint.

Answer

The endpoint is already InService, so it cannot be updated

Answer

The old model image is no longer available

Answer

The instance count is too low to deploy both variants

Question 6

A data scientist is training a binary classification model on a dataset with 100,000 positive samples and 1,000 negative samples. The model achieves 99% accuracy on the test set but a very low F1 score. What is the most likely cause?

Accepted Answer

The model predicts the majority class most of the time due to class imbalance. With 100,000 positive and 1,000 negative samples, the dataset is highly imbalanced (100:1 ratio). A model that simply predicts the majority class (positive) for all instances achieves 99% accuracy but fails to identify the minority class, resulting in a very low F1 score. This is the classic accuracy paradox in imbalanced classification.

Answer

The test set contains only positive samples

Answer

The model is overfitting due to too many features

Answer

The model is underfitting due to insufficient training

Question 7

A machine learning engineer needs to deploy a real-time inference endpoint for a model that requires GPU acceleration for low latency. Which AWS service should be used?

Accepted Answer

Amazon SageMaker real-time endpoint. Amazon SageMaker real-time endpoints are designed for low-latency, synchronous inference and support GPU instances for accelerated compute. They provide a persistent HTTPS endpoint that can be integrated with applications. Batch transform is for asynchronous, large-batch processing and does not provide a real-time endpoint. EC2 with auto scaling requires manual setup of inference server and scaling, lacking the managed experience. AWS Lambda does not support GPU acceleration.

Answer

Amazon SageMaker batch transform

Answer

Amazon EC2 with auto scaling

Answer

AWS Lambda with GPU

Question 8

A data scientist is training a deep learning model on Amazon SageMaker and notices that training is taking much longer than expected. The training job uses a single GPU instance. The model is a large transformer with millions of parameters. Which change would most likely reduce training time?

Accepted Answer

Use SageMaker's distributed data parallelism with multiple GPU instances. Distributed data parallelism in SageMaker splits the training dataset across multiple GPU instances, each computing gradients on its own shard and then synchronizing via all-reduce to update the model. This linearly scales throughput for large transformer models whose training time is dominated by forward and backward passes over massive datasets. Since the model already fits on a single GPU, data parallelism is the appropriate scaling strategy rather than model parallelism.

Answer

Reduce the batch size to fit in memory

Answer

Use a smaller instance type

Answer

Switch to a CPU instance

Question 9

A data scientist is building a regression model to predict housing prices. The dataset includes numerical features such as square footage, number of bedrooms, and year built, as well as categorical features such as neighborhood and roof type. Which TWO preprocessing steps are most important to apply before training a linear regression model?

Accepted Answer

One-hot encode categorical features. Option B is correct because linear regression requires numeric inputs, and categorical features such as neighborhood and roof type must be converted into a numeric representation; one-hot encoding creates binary indicator columns for each category, allowing the model to estimate a separate coefficient per category without imposing a false ordinal relationship. Option E is correct because the numerical features (square footage, number of bedrooms, year built) are on very different scales, and standardizing (z-score) or normalizing them improves the conditioning of the least-squares problem and ensures regularized or gradient-based training treats features comparably. Option A is not required here: PCA is a dimensionality-reduction technique that can help with multicollinearity or high dimensionality, but it is not an essential preprocessing step for this small, interpretable feature set and would sacrifice interpretability. Option C is not mandatory: outlier removal via the IQR rule can be useful for robustness, but it is a data-cleaning choice rather than a required preprocessing step and may discard legitimate high-price observations. Option D is incorrect because adding interaction terms between all features is not a necessary preprocessing step and would inflate the feature space, risking overfitting and multicollinearity without being specifically justified by the scenario.

Answer

Apply principal component analysis (PCA) for dimensionality reduction

Answer

Remove outliers using IQR

Answer

Add interaction terms between all features

Question 10

A machine learning engineer is evaluating a multi-class classification model that predicts product categories. The model outputs probabilities for 10 classes. The engineer wants to improve the model's calibration so that the predicted probabilities reflect the true likelihood of each class. Which THREE techniques can help?

Accepted Answer

Use temperature scaling. Temperature scaling (A) is correct because it divides the logits by a single learned scalar T before the softmax, directly adjusting the confidence of the 10-class probability outputs without changing the predicted class, making it the standard post-hoc calibration method for multi-class neural networks. Isotonic regression (B) is correct because it is a non-parametric calibration technique that fits a monotonic mapping from predicted probabilities to empirical frequencies, and it can be applied per-class or in a multiclass extension (e.g., one-vs-rest) to improve calibration. Platt scaling (D) is correct because it fits a logistic (sigmoid) transformation on the model's scores to produce calibrated probabilities, and although originally binary, it is commonly extended to multi-class via one-vs-rest or multiclass variants. Increasing model complexity (C) is not a calibration technique; it typically affects discrimination and can even worsen overconfidence rather than aligning probabilities with true likelihoods. Focal loss (E) is a training objective that reweights hard examples to address class imbalance, not a post-hoc or training-time calibration method for making predicted probabilities reflect true frequencies.

Answer

Increase model complexity

Answer

Use focal loss

Question 11

A company is using Amazon SageMaker to train a large language model with billions of parameters. The training job uses multiple GPU instances in a distributed fashion. The training is converging but the loss is not decreasing as expected. The data scientist suspects that the learning rate is too high. Which technique should the data scientist use to automatically adjust the learning rate during training?

Accepted Answer

Implement learning rate scheduling with a cosine annealing schedule. Cosine annealing is a learning rate schedule that starts with a relatively high learning rate and gradually decreases it following a cosine curve, which helps the model converge when the initial rate is too high. It automatically adjusts the learning rate during training without manual intervention, matching the requirement. This is a standard technique for large-model training where a fixed high rate causes the loss to plateau or oscillate.

Answer

Use a fixed learning rate and train for more epochs

Answer

Increase the batch size to reduce variance

Answer

Use gradient clipping

Question 12

A data scientist is building a recommendation system using collaborative filtering. The dataset contains user-item interactions in a sparse matrix. The model will be trained on Amazon SageMaker using the built-in Factorization Machines algorithm. Which data format should the scientist use for the training data?

Accepted Answer

RecordIO-protobuf format with sparse features. SageMaker's built-in Factorization Machines algorithm expects training data in RecordIO-protobuf format, which efficiently encodes sparse features common in collaborative filtering (user-item matrices). RecordIO-protobuf reduces storage and network overhead by storing only non-zero entries, which is critical when the interaction matrix is extremely sparse. CSV would bloat the payload with zeros and slow training.

Answer

CSV format with all features as columns

Answer

JSON format with nested arrays

Answer

Parquet format

Question 13

A machine learning engineer is deploying a model to an Amazon SageMaker endpoint for real-time inference. The model requires a preprocessing step that involves tokenizing text and converting it to a numerical format. To minimize latency, where should the preprocessing logic be implemented?

Accepted Answer

Inside the SageMaker inference container using the inference.py script. To minimize latency, it's best to include the preprocessing logic inside the inference container that serves the model. This avoids additional network calls to separate preprocessing services.

Answer

Using Amazon SageMaker batch transform

Answer

As a separate AWS Lambda function called before the endpoint

Answer

On the client side before sending the request

Question 14

A data scientist is training a deep learning model for image classification on Amazon SageMaker. The dataset consists of 10,000 images of size 224x224 pixels. The training job uses a single ml.p3.2xlarge instance. The data scientist notices that the GPU utilization is very low (~20%) and the training is slow. Which change would most likely improve GPU utilization?

Accepted Answer

Increase the number of data loader workers to load data in parallel. Low GPU utilization (~20%) during training typically indicates the GPU is starved for data — the CPU-bound data loading pipeline cannot feed batches fast enough. Increasing the number of data loader workers (D) parallelizes image decoding, augmentation, and batching across CPU cores, keeping the GPU busy. This is the classic fix for I/O-bound training on a single GPU instance.

Answer

Use gradient accumulation

Answer

Use a larger instance type with more GPUs

Answer

Increase the batch size

Question 15

A data scientist is training a binary classification model on an imbalanced dataset where the positive class represents 5% of the data. Which metric is most appropriate for evaluating model performance?

Accepted Answer

AUC-ROC. AUC-ROC evaluates the model's ability to rank positive instances above negatives across all classification thresholds, making it robust to class imbalance. With only 5% positives, accuracy is misleading because a trivial majority-class predictor achieves 95% accuracy. AUC-ROC provides a threshold-independent measure of discriminative power.

Answer

Accuracy

Answer

Root Mean Squared Error (RMSE)

Answer

R-squared