MLA-C01 Practice Test 35 — 15 Questions

Question 1

A data scientist is building a text classification model using a pre-trained BERT model from the Hugging Face library on SageMaker. The scientist wants to fine-tune the model on a custom dataset. Which TWO steps are necessary to set up the fine-tuning job? (Select TWO.)

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

Accepted Answer

Use the HuggingFace estimator provided by SageMaker. Option A is correct because the HuggingFace estimator from the SageMaker Python SDK is the purpose-built, supported interface for launching Hugging Face training jobs on SageMaker; it handles the training container, script entry point, and hyperparameters needed to fine-tune a pre-trained BERT model. Option D is correct because the HuggingFace estimator lets you pin the exact framework and library versions via the transformers_version and pytorch_version (or tensorflow_version) arguments, which is essential for reproducibility and compatibility with the pre-trained BERT checkpoint. Option B is not needed because SageMaker Clarify provides bias and explainability analysis, not a prerequisite for fine-tuning. Option C is unnecessary since the HuggingFace estimator already supplies a managed container with PyTorch and Transformers, so building a custom Docker image is only required for unsupported dependencies. Option E is not required because data preprocessing can be done inside the training script or beforehand; SageMaker Processing is optional and not a mandatory setup step for the fine-tuning job.

Answer

Enable SageMaker Clarify for explainability during training

Answer

Build a custom Docker container with PyTorch and Transformers

Answer

Use SageMaker Processing to preprocess the data in parallel

Question 2

A data scientist is building a text classification model using a pre-trained BERT model from the Hugging Face library on SageMaker. The scientist wants to fine-tune the model on a custom dataset. Which TWO steps are necessary to set up the fine-tuning job? (Select TWO.)

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

Accepted Answer

Convert the CSV file to the format of a single file with one instance per line.. Option C is correct because BlazingText expects the training data in a plain-text format where each line represents one instance, so the CSV must be converted to a single file with one instance per line. Option E is correct because BlazingText's text classification mode requires each line to contain a single text instance with its label prefixed by '__label__' (e.g., '__label__positive This movie was great'). Option A is incorrect because BlazingText performs its own tokenization and vectorization internally; one-hot encoding is not a required preprocessing step. Option B is incorrect because tokenization and stop-word removal are handled internally by the algorithm and are not required user steps. Option D is incorrect because uploading data to a notebook instance is not required; training data is read directly from Amazon S3.

Answer

Convert the text to one-hot encoded vectors.

Answer

Tokenize and remove stop words from the text.

Answer

Upload the data to an Amazon SageMaker notebook instance.

Question 3

Which TWO actions are recommended best practices when preparing training data for a machine learning model in AWS? (Choose two.)

Accepted Answer

Check for and handle missing values appropriately.. Option C is correct because missing values can bias or break training algorithms, so best practice is to detect them (e.g., with pandas isnull() or Amazon SageMaker Data Wrangler) and handle them via imputation, removal, or model-native handling. Option D is correct because splitting data into training, validation, and test sets lets you fit parameters, tune hyperparameters, and estimate generalization performance on unseen data, avoiding overfitting and data leakage. Option A is not recommended because not all outliers are errors; blindly removing them can discard legitimate signal and distort the distribution. Option B is wrong because training on the entire dataset leaves no held-out data for validation or unbiased testing. Option E is wrong because normalization is not always required and [0,1] scaling is only one of several techniques (e.g., standardization, log transforms) chosen based on the algorithm and feature distribution.

Answer

Remove all outliers from the dataset.

Answer

Train the model on the entire dataset to maximize data usage.

Answer

Always normalize all features to a [0,1] range.

Question 4

A data engineer is optimizing Amazon Athena queries on large datasets stored in S3 for machine learning data preparation. Which THREE practices improve query performance?

Accepted Answer

Partition the data by a frequently filtered column, such as date. Option A is correct because partitioning the S3 data by a frequently filtered column such as date lets Athena prune partitions and scan only the relevant prefixes, drastically reducing the amount of data read and thus improving performance and lowering cost. Option D is correct because columnar formats like Parquet or ORC allow Athena to read only the columns referenced by a query and provide better compression and predicate pushdown, minimizing I/O compared to row-based formats. Option E is correct because compressing data with Snappy or gzip reduces the bytes transferred from S3 and read by Athena, and these splittable/columnar-friendly codecs work well with Parquet/ORC to further speed up scans. Option B is incorrect because uncompressed CSV is a row-based, non-splittable-friendly format that forces Athena to scan all columns and more bytes, hurting performance. Option C is incorrect because partitioning by every column creates an excessive number of small partitions and files, which increases metadata overhead and can degrade rather than improve query performance.

Answer

Use uncompressed CSV files for simplicity

Answer

Partition the data by every column to maximize filtering

Question 5

A company uses Amazon SageMaker to deploy a model for real-time inference. They want to perform A/B testing between two model versions. Which TWO actions should the company take to set up A/B testing? (Choose TWO.)

Accepted Answer

Create an endpoint configuration with multiple production variants, each with a different model.. Option A is correct because SageMaker A/B testing is implemented by creating an endpoint configuration that contains multiple production variants, where each variant references a different model (via its ModelName), allowing the endpoint to serve both model versions behind a single endpoint. Option C is correct because each production variant has an InitialVariantWeight, and setting these weights establishes the desired traffic split (e.g., 80/20) so the endpoint routes the corresponding proportion of invocations to each model. Option B is incorrect because CloudWatch Evidently is a feature-flagging/experimentation service for applications, not the mechanism for splitting traffic between SageMaker production variants. Option D is incorrect because auto scaling adjusts instance counts for capacity, not traffic distribution between model versions, so it does not set up A/B testing. Option E is incorrect because setting a variant's weight to 0 means it receives no traffic, and shifting to 100 later is a blue/green-style cutover rather than a concurrent A/B test with a defined split.

Answer

Use Amazon CloudWatch Evidently to split traffic between models.

Answer

Enable auto scaling for each production variant individually.

Answer

Set the second production variant's weight to 0 and update later to 100.

Question 6

A healthcare startup has deployed a machine learning model on Amazon SageMaker that predicts patient readmission risks. The model uses sensitive health data stored in an S3 bucket encrypted with AWS KMS. The SageMaker endpoint is configured with an IAM role that has the following policy attached: {    "Version": "2012-10-17",    "Statement": [        {            "Effect": "Allow",            "Action": "s3:*",            "Resource": "arn:aws:s3:::healthcare-data/*",            "Condition": {                "Bool": {                    "aws:SecureTransport": "true"                }            }        },        {            "Effect": "Allow",            "Action": "kms:Decrypt",            "Resource": "*"        }    ] }. During a security audit, the team discovers that the IAM role's KMS permission is too permissive because it allows decryption of any KMS key in the account. The team needs to modify the policy to follow the principle of least privilege while still allowing the SageMaker endpoint to read the encrypted data. Which modification should the team make?

Accepted Answer

Change the KMS statement to: "Action": "kms:Decrypt", "Resource": "arn:aws:kms:us-east-1:123456789012:key/1234abcd-12ab-34cd-56ef-1234567890ab". The KMS statement currently allows kms:Decrypt on Resource '*' , which permits decryption with any KMS key in the account. Restricting the Resource to the specific KMS key ARN used to encrypt the S3 bucket enforces least privilege while still allowing the SageMaker endpoint to decrypt the data. This is the correct modification to scope the permission to only the necessary key.

Answer

Change the KMS statement Action to "kms:DescribeKey" instead of "kms:Decrypt"

Answer

Add a condition to the KMS statement: "Condition": { "StringEquals": { "kms:ViaService": "s3.us-east-1.amazonaws.com" } }

Answer

Remove the KMS statement entirely, as S3 bucket policies with SSE-KMS do not require KMS permissions

Question 7

A company deploys a real-time inference endpoint with auto-scaling using a target tracking policy based on average Invocations per instance. They notice that during a traffic spike, the endpoint scales out too late, causing increased latency. They want to scale proactively before the spike. Which strategy should they implement?

Accepted Answer

Use a scheduled scaling action to add capacity before the expected spike. A scheduled scaling action adds capacity at a predetermined time before the expected traffic spike, allowing the endpoint to be ready proactively. Target tracking reacts after metrics breach thresholds, which is inherently reactive and too slow for sharp spikes. Scheduled scaling is the correct strategy when spikes are predictable (e.g., known business hours or events).

Answer

Enable provisioned concurrency on the endpoint

Answer

Pre-warm the endpoint by sending dummy requests

Answer

Switch to a step scaling policy with a higher cooldown period

Question 8

A machine learning engineer is preparing a dataset for a binary classification model. The dataset has 10,000 samples with a 1:100 class imbalance. The engineer needs to balance the classes before training. Which technique would create a balanced dataset without discarding majority class samples and without generating synthetic data?

Accepted Answer

Random oversampling of the minority class. Random oversampling of the minority class duplicates existing minority samples until the class distribution is balanced, which increases the minority count without discarding any majority samples and without creating synthetic data. This satisfies all three constraints: balanced dataset, no majority samples removed, and no synthetic generation. It is a straightforward resampling technique commonly used as a baseline before considering more advanced methods.

Answer

Cost-sensitive learning with class weights

Answer

Synthetic Minority Over-sampling Technique (SMOTE)

Answer

Random undersampling of the majority class

Question 9

A data scientist wants to train a model on SageMaker using a custom PyTorch script, then register the best model in the SageMaker Model Registry. The training job is part of a SageMaker Pipeline. Which pipeline step should be used to register the model?

Accepted Answer

RegisterModelStep. The `RegisterModelStep` is specifically designed to create a model resource and register it in the SageMaker Model Registry as part of a pipeline. It takes the training output (e.g., model artifacts from a `TrainingStep`) and packages it with the specified inference image and metadata, then creates a model package group version. This is the correct step for registering a model after training, as it directly integrates with the Model Registry for versioning and approval workflows.

Answer

CreateModelStep

Answer

TrainingStep

Answer

TransformStep

Question 10

A company is fine-tuning a large language model using reinforcement learning from human feedback (RLHF). Which THREE components are typically required?

Accepted Answer

A reference model. In the standard RLHF pipeline, the policy model (D) is the large language model being fine-tuned; it generates responses and is updated via reinforcement learning (typically PPO) to maximize reward. A reward model (C) is trained on human preference comparisons to output a scalar score predicting human preference, and it supplies the reward signal that guides the policy's optimization. A reference model (B) is a frozen copy of the pre-RLHF model used to compute a KL-divergence penalty, keeping the policy from drifting too far from the original model and collapsing into degenerate, high-reward outputs. The other options are not required components: a discriminative classifier (A) is not part of RLHF (the reward model is a regression-style preference predictor, not a classifier), and a value function (E) is only an internal component of the PPO algorithm used to estimate advantages, not a standalone required model in the RLHF architecture.

Answer

A discriminative classifier

Answer

A value function

Question 11

A team is training a PyTorch model using SageMaker and wants to use their own custom training container with a specific PyTorch version. Which approach should they use?

Accepted Answer

Use SageMaker Bring Your Own Container (BYOC) with a custom Docker image. When a team needs a custom training container with a specific PyTorch version or custom dependencies that are not available in the built-in SageMaker images, the correct approach is Bring Your Own Container (BYOC), where they build a Docker image and push it to Amazon ECR, then reference it in the SageMaker estimator. This gives full control over the framework version and environment.

Answer

Use the SageMaker built-in PyTorch estimator and set the framework_version

Answer

Use SageMaker Script Mode with a PyTorch script

Answer

Use SageMaker Autopilot to automatically select the container

Question 12

A team is fine-tuning a large language model (LLM) using SageMaker and wants to reduce memory footprint during training. Which technique should they use?

Accepted Answer

Use QLoRA (Quantized Low-Rank Adaptation) with 4-bit quantization. QLoRA combines 4-bit quantization of the base model weights (via NF4) with LoRA adapters, dramatically reducing GPU memory during fine-tuning while preserving accuracy close to full fine-tuning. The 4-bit quantization is the key memory-saving mechanism, since the frozen base weights consume ~4x less memory than fp16/fp32. LoRA alone reduces optimizer/gradient memory but still loads the base model in higher precision, so QLoRA is the correct answer for minimizing memory footprint.

Answer

Use LoRA (Low-Rank Adaptation) with fp32 precision

Answer

Use SageMaker Model Parallelism with tensor parallelism

Answer

Full fine-tuning on a p3.16xlarge instance

Question 13

A data scientist needs to label a large dataset of product images for a classification model. They want to reduce labeling costs by prioritizing uncertain samples. Which Amazon SageMaker Ground Truth feature should they use?

Accepted Answer

Active learning. Active learning in Amazon SageMaker Ground Truth automatically selects the most uncertain or informative samples from the unlabeled dataset and sends them to human annotators. This prioritization reduces labeling costs by focusing budget on the samples that will most improve model performance, rather than labeling all data indiscriminately.

Answer

Pre-built worker templates

Answer

Consolidated labeling

Answer

Automated data labeling

Question 14

Which SageMaker built-in algorithm is best suited for detecting anomalous login attempts based on IP addresses and user behavior?

Accepted Answer

IP Insights. IP Insights is a SageMaker built-in algorithm purpose-built for learning the patterns of IP address usage by users and resources, making it ideal for detecting anomalous login attempts. It uses a neural network to embed IP addresses and entities (like user IDs) into a vector space, flagging unusual combinations as anomalies. XGBoost, PCA, and K-Means are general-purpose algorithms not designed for IP-entity behavioral analysis.

Answer

XGBoost

Answer

PCA

Answer

K-Means

Question 15

A machine learning engineer is using Amazon SageMaker Debugger to monitor a training job for a deep neural network. They receive a rule alert indicating 'exploding gradients'. Which action should they take to address this issue?

Accepted Answer

Reduce the learning rate. Exploding gradients occur when large error gradients accumulate during backpropagation, causing unstable updates and divergence. Reducing the learning rate directly scales down the parameter update step (Δθ = -η∇J), preventing the weights from overshooting and stabilizing training. This is the standard first-line remedy for exploding gradients in deep networks.

Answer

Use a smaller batch size

Answer

Increase the number of layers to absorb gradients

Answer

Increase the learning rate