Skip to content

Skill 1.3.2: Label and annotate data

Question 1 - SageMaker Ground Truth Labeling Workflow Design

A healthcare AI startup is building a model to detect abnormal chest X-ray patterns using Amazon SageMaker AI. They have 50,000 unlabeled X-ray images in an S3 bucket and need high-quality annotations from certified radiologists.

The compliance team mandates that all annotation work remains within a private VPC with no public internet access, and the ML engineering team wants to minimize operational overhead for the labeling infrastructure.

Which approach should the engineering team use to label this dataset? (Select ONE)

  • A) Deploy an open-source annotation tool (e.g., Label Studio) on EC2 instances within the VPC and configure security groups to restrict access to radiologists.
  • B) Use SageMaker Ground Truth with a private workforce using the Amazon Cognito user pool and a VPC endpoint for secure access.
  • C) Create a custom AWS Lambda function that serves images to radiologists via an API Gateway private endpoint, with annotations stored in DynamoDB.
  • D) Use Amazon Bedrock to automatically generate annotations from the X-ray images without human review.
  • E) Send the images to Amazon Mechanical Turk and filter workers by their self-reported medical qualifications.
Click to reveal Answer & Explanation

Correct Answer: B

Correct Option Analysis

  • B is Correct: SageMaker Ground Truth natively supports private workforces via Amazon Cognito user pools, allowing organizations to restrict annotation tasks to authenticated internal users (certified radiologists). VPC endpoints for Ground Truth ensure all traffic stays within the private network, satisfying compliance requirements with minimal operational overhead since Ground Truth manages the labeling infrastructure, task distribution, and result collection.

Distractor Analysis

  • A is Incorrect: Deploying and maintaining a custom annotation tool on EC2 requires patching, scaling, high availability configuration, and ongoing operational overhead. This violates the "minimize operational overhead" constraint when SageMaker Ground Truth provides this natively.
  • C is Incorrect: Building a custom annotation application using Lambda, API Gateway, and DynamoDB introduces significant development and maintenance overhead. Ground Truth already provides these capabilities as a managed service.
  • D is Incorrect: Bedrock cannot automatically generate reliable medical image annotations that meet the accuracy requirements of certified radiologists. Medical imaging annotation requires domain expertise that cannot be fully automated, and compliance standards would likely reject purely AI-generated labels.
  • E is Incorrect: Amazon Mechanical Turk is a public marketplace workforce. Workers cannot be reliably verified as certified radiologists, which violates the compliance requirement for certified professionals handling patient data.

Key Takeaway: For enterprise labeling workflows requiring private, authenticated workforces and minimal operational overhead, SageMaker Ground Truth with private workforce and VPC endpoints is the native AWS solution.

Question 2 - Labeling Instruction Ambiguity Resolution

A financial services company is building a document classification model using SageMaker Ground Truth to route customer support tickets. The annotation team has labeled 10,000 support tickets across 5 complaint categories. After reviewing quality metrics, the ML engineer discovers that two annotators achieve high agreement on most categories but consistently disagree when classifying tickets about "payment disputes" versus "billing errors."

The model's errors cluster specifically at this boundary, resulting in misclassified tickets that are routed to the wrong department.

What is the MOST appropriate action to resolve this labeling quality issue? (Select ONE)

  • A) Add more annotators to the task to increase annotation throughput.
  • B) Remove the "billing errors" category entirely since it overlaps with "payment disputes."
  • C) Review and refine the labeling instructions with clear, mutually exclusive definitions for the two categories, including edge-case examples, then re-label the disputed tickets.
  • D) Use AWS Glue DataBrew to automatically merge the two categories into one combined class.
  • E) Implement a majority vote using SageMaker Ground Truth built-in annotation consolidation and accept the current labels.
Click to reveal Answer & Explanation

Correct Answer: C

Correct Option Analysis

  • C is Correct: Persistent inter-annotator disagreement at a specific class boundary signals ambiguous labeling instructions, not annotator error. Refining the instructions with precise, mutually exclusive definitions and providing edge-case examples directly addresses the root cause. Re-labeling the disputed tickets with the clarified guidelines improves label quality, which sets the ceiling for model performance.

Distractor Analysis

  • A is Incorrect: Adding more annotators only adds more inconsistent labels if the instructions remain ambiguous. It treats the symptom (disagreement) rather than the root cause (unclear guidelines).
  • B is Incorrect: Merging legitimate business categories to avoid ambiguity may distort business processes. Payment disputes and billing errors likely require different routing destinations, so separating them is necessary for correct ticket handling.
  • D is Incorrect: AWS Glue DataBrew is a data profiling and cleaning tool for tabular data, not a document classification label resolver. It cannot intelligently decide class boundaries or merge semantic categories.
  • E is Incorrect: Majority vote consolidation with the existing ambiguous labels merely averages out the disagreement without resolving the root ambiguity. The model will continue to learn a blurred boundary and produce errors at the same edge cases.

Key Takeaway: When annotators consistently disagree on specific class boundaries, the labeling instructions—not the annotators—are the root cause; clarify definitions and edge-case examples before relabeling.

Question 3 - Multi-Modal Dataset Bias Mitigation

An e-commerce platform is training a multimodal product recommendation model using product images, text descriptions, and numeric attributes (price, rating, category). The dataset was collected over 12 months, but due to a marketing initiative, 80% of product reviews come from Q4, and one product category (electronics) appears in only 2% of samples.

The ML engineer needs to prepare the dataset for training to prevent the model from overfitting to seasonal patterns while ensuring the underrepresented category is adequately represented.

Which TWO techniques should the engineer apply? (Select TWO)

  • A) Shuffle the dataset randomly before splitting to break temporal correlation between collection order and training batches.
  • B) Apply time-based splitting (e.g., train on months 1-9, validate on months 10-12) to respect chronological order.
  • C) Apply data augmentation specifically to the electronics category (e.g., synthetic image variations, text paraphrasing) to increase its representation.
  • D) Apply SMOTE oversampling uniformly across all categories to balance the dataset.
  • E) Use stratified splitting to ensure the electronics category appears proportionally in both training and validation sets.
Click to reveal Answer & Explanation

Correct Answers: A and C

Correct Options Analysis

  • A is Correct: When data is collected in batches with temporal or seasonal correlation (e.g., heavy Q4 representation), shuffling breaks the alignment between collection order and mini-batch sequence, ensuring each training batch contains a representative mix of time periods rather than being dominated by Q4 records.
  • C is Correct: Data augmentation (image transformations, text paraphrasing, synthetic variations) adds diversity to the underrepresented electronics category without duplicating identical records, which could cause overfitting to that specific narrow slice.

Distractor Analysis

  • B is Incorrect: Time-based splitting is appropriate for time-series forecasting where chronological order matters, but for product recommendation the goal is to generalize across all time periods, making random shuffling and stratified splits more suitable than rigid chronological splits.
  • D is Incorrect: SMOTE applied uniformly across all categories would balance the dataset indiscriminately, potentially overrepresenting the majority categories already well-covered and diluting the specific focus needed on the electronics category.
  • E is Incorrect: While stratified splitting is a valuable technique for ensuring representation in evaluation sets, it does not address the underlying underrepresentation within the training data itself. The 2% representation needs augmentation or oversampling to improve the model's learning signal.

Key Takeaway: Match the technique to the specific defect: shuffling fixes temporal ordering biases, stratification ensures balanced evaluation splits, and targeted augmentation addresses underrepresentation for specific groups.

Question 4 - Class Imbalance in Text Classification

A fraud detection team is building a model to classify financial transactions as "fraudulent" or "legitimate" using a dataset of 1 million transactions, where only 0.5% are fraudulent. A baseline model trained on this data achieves 99.5% accuracy but detects only 12% of actual fraudulent transactions.

The team needs to improve the model's ability to detect fraud (recall) while maintaining reasonable precision (avoiding excessive false alerts).

Which TWO techniques should the ML engineer apply to resolve the class imbalance? (Select TWO)

  • A) Undersample the majority class (legitimate transactions) to balance the dataset, accepting some loss of information from discarded majority samples.
  • B) Oversample the minority class (fraudulent transactions) using SMOTE to generate synthetic fraud examples with feature space interpolation.
  • C) Report overall accuracy as the primary evaluation metric, since it already shows 99.5% performance.
  • D) Use class weighting in the loss function to penalize misclassifications of fraudulent transactions more heavily during training.
  • E) Remove all majority class samples that the model correctly classifies, since they add no information to the training process.
Click to reveal Answer & Explanation

Correct Answers: A and D

Correct Options Analysis

  • A is Correct: Undersampling the majority class (legitimate transactions) reduces the imbalance and prevents the model from being dominated by the common label. While it discards potentially informative samples, with 1 million records and only 5,000 fraud cases, a balanced undersampled dataset still provides sufficient legitimate examples for training.
  • D is Correct: Class weighting adjusts the loss function proportionally to class frequency (or a chosen ratio), penalizing misclassified fraud cases more heavily. This helps the model prioritize minority class detection without needing to modify the dataset itself.

Distractor Analysis

  • B is Incorrect: SMOTE is generally not recommended for high-dimensional and mixed-type tabular transaction data, especially with highly imbalanced fraud patterns. Synthetic interpolation between fraud examples may generate unrealistic feature combinations that do not represent actual fraudulent behavior and could degrade precision rather than improve recall.
  • C is Incorrect: Accuracy is misleading in imbalanced datasets, as demonstrated by the 99.5% accuracy with only 12% fraud detection. The team should report metrics like precision, recall, F1-score, and PR-AUC that specifically measure minority class performance.
  • E is Incorrect: Removing correctly classified majority samples is a specific form of undersampling, but selecting "correctly classified" samples introduces bias by removing only "easy" legitimate cases, potentially overrepresenting difficult legitimate cases that share features with fraud and confusing the model further.

Key Takeaway: For class imbalance, report precision-recall metrics, not overall accuracy, and apply rebalancing techniques like undersampling, class weighting, or targeted collection of rare-class data.

Question 5 - SageMaker Ground Truth Automated Data Labeling with Active Learning

A computer vision startup is building a model to detect manufacturing defects in automotive parts using SageMaker AI. They have 200,000 unlabeled images and a limited annotation budget. The team wants to use SageMaker Ground Truth with automated data labeling.

What is the correct workflow to maximize label quality while minimizing total annotation cost? (Select ONE)

  • A) Use Ground Truth automated data labeling with a small human-reviewed labeling job first, then let the auto-labeling feature label remaining images using confidence thresholds; route low-confidence samples to human review.
  • B) Launch an automated data labeling job that auto-labels all 200,000 images without any initial human-labeled dataset.
  • C) Send all images to a public Mechanical Turk workforce to maximize throughput.
  • D) Use Ground Truth with a private workforce to manually label all 200,000 images, since automated labeling cannot match human accuracy.
  • E) Skip Ground Truth and use Amazon Rekognition Custom Labels to auto-generate labels without human intervention.
Click to reveal Answer & Explanation

Correct Answer: A

Correct Option Analysis

  • A is Correct: SageMaker Ground Truth automated data labeling uses active learning: it starts with a small human-labeled dataset, trains an intermediate model, and auto-labels only images where the model has high confidence, routing low-confidence images to human workers. This optimizes the human labeling budget while maintaining label quality.

Distractor Analysis

  • B is Incorrect: Ground Truth automated data labeling requires an initial human-labeled dataset to bootstrap the active learning loop. Without any human-labeled data, the model cannot learn to make reliable predictions.
  • C is Incorrect: Using Mechanical Turk for 200,000 images would exceed the annotation budget and may not provide the manufacturing domain expertise needed for accurate defect detection. Ground Truth automated labeling reduces human workload on easy cases.
  • D is Incorrect: Manually labeling all 200,000 images is operationally expensive, time-consuming, and unnecessary when automated data labeling can handle high-confidence cases reliably. The question specifies a limited budget.
  • E is Incorrect: Amazon Rekognition Custom Labels requires labeled training data to learn new custom labels. It cannot automatically generate high-quality labels for a novel domain like manufacturing defects without a human-labeled training set.

Key Takeaway: Ground Truth automated labeling reduces annotation cost by bootstrapping from a small human-labeled set and only routing low-confidence samples to human reviewers, applying active learning.

Question 6 - Label Data Type and Annotation Storage Architecture

A data engineering team is building a document processing pipeline that extracts entities from legal contracts using Amazon Comprehend and Amazon Textract. A team of legal experts needs to review and annotate the extracted entities for correctness before the data is used to fine-tune a custom entity recognition model.

The annotated data must be stored in a format that supports both human review and downstream model fine-tuning in SageMaker AI.

Which AWS architecture should the team implement for the annotation workflow? (Select ONE)

  • A) Store the original contracts in S3. Use SageMaker Ground Truth with a private workforce of legal experts to review and correct entity annotations. Store the augmented manifest in S3 for downstream fine-tuning.
  • B) Extract entities using Comprehend and store annotations as CSV files in an RDS database, with a custom web app for legal expert review.
  • C) Use SageMaker Data Wrangler to build a data flow that extracts entities and exports them directly to SageMaker Feature Store for model training.
  • D) Store annotated entities in Amazon DynamoDB with a Lambda trigger to validate annotation quality before writing to S3.
  • E) Use AWS Glue jobs to transform extracted entities into Parquet format and skip human review since Comprehend provides high accuracy.
Click to reveal Answer & Explanation

Correct Answer: A

Correct Option Analysis

  • A is Correct: SageMaker Ground Truth supports a native entity annotation workflow where human reviewers can validate and correct extracted entity boundaries from documents. The augmented manifest (JSON Lines) stores both original data and annotation labels in a format compatible with SageMaker fine-tuning jobs, with S3 as the durable storage.

Distractor Analysis

  • B is Incorrect: Building a custom web application for annotation introduces unnecessary operational overhead. Ground Truth already provides a managed annotation UI, workflow management, and result formats that integrate with SageMaker training.
  • C is Incorrect: SageMaker Data Wrangler is for feature transformation and data preparation, not for human annotation workflows. It does not provide a mechanism for legal experts to review and correct entity annotations.
  • D is Incorrect: Using DynamoDB and Lambda for annotation validation adds operational complexity (custom validation code, database schema design, versioning) without providing benefit over Ground Truth's managed workflow, quality metrics, and annotation consolidation.
  • E is Incorrect: Skipping human review and relying solely on Comprehend's automated outputs ignores the requirement for human expert correction. Comprehend's accuracy in complex legal documents may not meet the quality bar for fine-tuning without human validation.

Key Takeaway: SageMaker Ground Truth provides native annotation workflows and augmented manifest formats that integrate directly with fine-tuning jobs, eliminating custom infrastructure for human-in-the-loop labeling.

Question 7 - Prompt-Response Pair Validation for Generative AI Fine-tuning

A startup is fine-tuning a foundation model (via Amazon Bedrock or SageMaker JumpStart) using a custom dataset of customer support conversations. The dataset contains 15,000 prompt-response pairs scraped from historical chat logs. Before submitting the fine-tuning job, the ML engineer must validate the training data integrity.

What are TWO mandatory validation checks the engineer should implement? (Select TWO)

  • A) Verify that each response is relevant to its prompt and that no prompt-response pair has missing or malformed fields.
  • B) Run content safety screening (e.g., Guardrails) to detect PII, toxicity, or inappropriate content in both prompts and responses before fine-tuning.
  • C) Apply data augmentation to the prompt-response pairs to increase dataset size before validation.
  • D) Shuffle the prompt-response pairs to remove any sequential ordering bias.
  • E) Use Amazon Comprehend to translate all prompts and responses to English before fine-tuning.
Click to reveal Answer & Explanation

Correct Answers: A and B

Correct Options Analysis

  • A is Correct: Structural data integrity requires verifying that each prompt has a corresponding response, no fields are missing or truncated, and the response semantically answers its prompt. Malformed pairs (e.g., empty responses, mismatched prompt-response) teach the model incorrect associations.
  • B is Correct: For generative AI fine-tuning, safety screening is critical before customization because unsafe content (PII, toxic language, offensive material) that enters the fine-tuning dataset can be reproduced in model outputs. Amazon Bedrock Guardrails or SageMaker content classifiers can automate this screening.

Distractor Analysis

  • C is Incorrect: Augmentation should occur after validation, not before. Validating a dataset after augmenting it means the synthetic examples may not have undergone the same integrity checks, allowing malformed synthetic data into the fine-tuning set.
  • D is Incorrect: While shuffling is useful for training batch randomization, it does not validate the structural integrity of each prompt-response pair. Shuffling reordered but still invalid pairs would not fix malformed content.
  • E is Incorrect: Amazon Comprehend supports translation only between English and a limited set of primary languages, and translating a multilingual customer support dataset to English would change the domain's semantic characteristics. If the model must serve multilingual customers, retaining the original languages is correct; validation should be done on the original data.

Key Takeaway: For fine-tuning generative AI models, integrity checks must verify structural completeness (well-formed prompt-response pairs) and content safety (no PII/toxic content) before data reaches model customization.

Question 8 - AWS Glue Data Quality Rule-based Validation for ML Datasets

A marketing analytics team uses AWS Glue ETL jobs to prepare a customer segmentation dataset for ML training. The dataset is loaded into an S3 bucket as Parquet files and registered in the AWS Glue Data Catalog. Before training, the team must ensure that each customer record has all required fields, values are within expected ranges (e.g., age 18–100), and duplicate customer IDs are deleted.

The team wants to automate this validation with repeatable rules that fail the pipeline if violations exceed acceptable thresholds.

Which AWS service should the team use, and what is the correct approach? (Select ONE)

  • A) Use AWS Glue Data Quality with a Data Quality Definition Language (DQDL) ruleset that checks completeness, column value ranges, and uniqueness constraints, then evaluate the ruleset on each Glue job run.
  • B) Use AWS Glue DataBrew profile jobs to visually inspect data quality and manually review the results after each pipeline run.
  • C) Use Amazon SageMaker Data Wrangler to write custom Python code that validates data in real-time during feature transformation.
  • D) Use Amazon QuickSight to visualize duplicate customer IDs and identify invalid ranges before manually correcting the dataset.
  • E) Use AWS Lambda functions triggered by S3 events to run validation queries and send notifications when errors are detected.
Click to reveal Answer & Explanation

Correct Answer: A

Correct Option Analysis

  • A is Correct: AWS Glue Data Quality is purpose-built for declaring explicit DQDL rules (e.g., IsComplete "customer_id", ColumnValues "age" between 18 and 100, ColumnStatistics "customer_id" distinctCount == rowCount) and evaluating them automatically on each Glue job run. Rulesets are repeatable, auditable, and can fail the pipeline when thresholds are exceeded.

Distractor Analysis

  • B is Incorrect: DataBrew profile jobs provide automated data profiling but are not designed for rule-based validation that fails pipelines. They output visual summaries that require manual review, which does not meet the automation requirement.
  • C is Incorrect: Data Wrangler is an interactive data transformation tool for building feature engineering pipelines, not a declarative rule-based validation service. Writing custom Python validation logic adds unnecessary code complexity that Glue Data Quality already handles natively.
  • D is Incorrect: QuickSight is a BI visualization tool, not a data validation pipeline. Visual inspection of duplicates is not repeatable, does not fail the pipeline automatically, and requires manual intervention for every pipeline run.
  • E is Incorrect: Writing custom Lambda validation introduces operational overhead in script maintenance, error handling, and scaling. AWS Glue Data Quality provides declarative rulesets with built-in threshold evaluation and integrates with Glue ETL natively.

Key Takeaway: Glue Data Quality uses DQDL rulesets to declaratively enforce completeness, value ranges, uniqueness, and other quality constraints on every job run without custom validation code.

Question 9 - Data Deduplication Strategy for Enterprise Knowledge Base

An AI startup is preparing a dataset for fine-tuning a specialized text generation model using historical customer support conversations from multiple sources (Zendesk, Slack, email). The team discovers that some support queries appear multiple times across different sources with slight variations (e.g., "I need to reset my password" vs. "password reset requested"). Some are exact duplicates, while others are near-duplicates with minor wording differences.

The ML engineer needs to reduce redundancy while preserving legitimate variations of similar topics. What is the MOST appropriate deduplication approach? (Select ONE)

  • A) Apply exact string matching to remove all records with identical text content, then manually review remaining near-duplicates to determine uniqueness based on feature-level differences.
  • B) Use AWS Glue DataBrew to automatically detect and remove exact and near-duplicate records using its built-in deduplication feature, reviewing auto-suggested thresholds.
  • C) Remove all records that share common keywords (e.g., "password reset"), keeping only one representative record per keyword.
  • D) Use AWS Glue ETL jobs to hash each record and remove records with identical hash values, ignoring all near-duplicates.
  • E) Delete all records from any source except the primary source (e.g., Zendesk) since other sources are redundant.
Click to reveal Answer & Explanation

Correct Answer: B

Correct Option Analysis

  • B is Correct: AWS Glue DataBrew provides a built-in deduplication feature that can detect both exact duplicates and fuzzy/near-duplicates based on configurable matching thresholds (e.g., Levenshtein edit distance). This allows the engineer to remove redundant records while preserving legitimate variations that carry distinct information, with the flexibility to review and adjust the threshold.

Distractor Analysis

  • A is Incorrect: Exact string matching only removes identical records but leaves all near-duplicates. Manual review of thousands of near-duplicate records is impractical for large datasets and operational overhead.
  • C is Incorrect: Keyword-based removal is overly aggressive: legitimate variations of the same topic may carry distinct context (e.g., different error scenarios for password reset) that the model needs to learn. This approach would collapse valuable distinctions and reduce dataset diversity.
  • D is Incorrect: Hash-based exact deduplication has the same limitation as exact string matching: it only identifies records with identical content. Near-duplicates with slight wording differences will have different hash values and remain in the dataset.
  • E is Incorrect: Removing entire data sources violates the requirement to preserve legitimate variations. Different communication channels may capture different facets of the same underlying issues (e.g., email may contain long-form context while Slack has brief exchanges), and deleting them reduces the dataset's representational richness.

Key Takeaway: For deduplication, use DataBrew's fuzzy matching capabilities to distinguish true duplicates from legitimate near-duplicate variations with context-dependent differences, rather than relying on exact matching or keyword heuristics.

Question 10 - Inter-annotator Agreement Metrics and Ground Truth Quality Metrics

A legal tech company is building a document classification model using SageMaker Ground Truth to categorize contracts into 10 types (e.g., NDA, MSA, SOW, etc.). The annotation team consists of 5 legal experts. The ML engineer needs to measure the consistency of annotations to determine if the labeling quality meets the standard for model training.

Which metrics should the engineer use in SageMaker Ground Truth to evaluate annotation quality? (Select TWO)

  • A) Fleiss's kappa or Cohen's kappa to measure inter-annotator agreement across all 5 annotators on multi-class labels.
  • B) Ground Truth's built-in annotation consolidation metrics, including label confidence scores, to identify low-agreement items for review.
  • C) Overall model accuracy after training on the labeled data to assess annotation quality.
  • D) The total number of labels produced per hour per annotator as a proxy for quality.
  • E) Cross-entropy loss of the trained model on the validation split to measure label correctness.
Click to reveal Answer & Explanation

Correct Answers: A and B

Correct Options Analysis

  • A is Correct: Kappa statistics (Fleiss's kappa for multiple annotators, Cohen's kappa for pairwise agreement) measure the degree of agreement beyond chance among annotators. High kappa values (typically >0.7) indicate reliable annotations, while low values suggest ambiguous instructions or label noise.
  • B is Correct: SageMaker Ground Truth provides built-in annotation consolidation with confidence scores and quality metrics that show which data items have high disagreement among annotators. Engineers can route low-confidence items to additional review or inspect them for instruction ambiguity.

Distractor Analysis

  • C is Incorrect: Overall model accuracy after training is a downstream performance metric, not an annotation quality measure. A model can achieve high accuracy despite poor annotation quality (e.g., by overfitting to noise or learning a majority-class predictor), or low accuracy despite high-quality labels (due to modeling issues).
  • D is Incorrect: Annotator throughput (labels per hour) measures efficiency, not quality. Fast labeling may correlate with errors, while slow labeling may indicate thoroughness or difficulty. Quality should be measured by consistency and correctness, not speed.
  • E is Incorrect: Cross-entropy loss during model training measures how well the model fits the labeled data, not whether the labels themselves are correct. If the labels are noisy, the model may struggle to reduce loss, but low loss does not confirm label quality—it may simply mean the model overfits to noise.

Key Takeaway: Label quality is best assessed using inter-annotator agreement metrics (e.g., kappa) and Ground Truth's built-in confidence scores, which flag low-agreement items for review before model training begins.

Question 11 - Labeling Workflow for Streaming Data with Amazon Kinesis

A logistics company is building a real-time route optimization model that classifies incoming GPS telemetry data from delivery vehicles. The dataset arrives as a continuous stream through Amazon Kinesis Data Streams. A team of 3 dispatchers must review and annotate a subset of the streaming events (e.g., labeling detour events as "traffic-related", "accident-related", "weather-related") to create a labeled training dataset.

The ML engineer must design a labeling pipeline that minimizes operational overhead and integrates with the streaming architecture.

Which TWO services should the engineer include in the architecture? (Select TWO)

  • A) Use Amazon Kinesis Data Firehose to batch streaming events into Amazon S3, then launch a SageMaker Ground Truth labeling job on the batched data using a private workforce.
  • B) Use AWS Lambda to write custom labeling logic that automatically classifies each streaming event as it arrives, without human review.
  • C) Use Amazon SageMaker Data Wrangler to analyze streaming data in real-time and apply labels through its built-in annotation features.
  • D) Use AWS Glue to load streaming events into a Redshift cluster, then deploy a custom application for dispatchers to label events.
  • E) Use Amazon Managed Streaming for Apache Kafka (MSK) to buffer streaming events and trigger a SageMaker Ground Truth real-time labeling workflow.
Click to reveal Answer & Explanation

Correct Answer: A

Correct Option Analysis

  • A is Correct: Kinesis Data Firehose natively batches and delivers streaming data to S3, creating the durable dataset needed for a Ground Truth labeling job. Ground Truth supports private workforces via Cognito and manages the labeling workflow, annotation UI, and result collection with minimal operational overhead.

Distractor Analysis

  • B is Incorrect: Lambda functions cannot replace human expertise for annotating complex detour events. The requirement explicitly states that dispatchers must review and annotate the events, making human-in-the-loop annotation necessary.
  • C is Incorrect: SageMaker Data Wrangler is an interactive data transformation tool for batch feature engineering, not a real-time streaming annotation service. It does not support native annotation workflow or integration with Kinesis streams.
  • D is Incorrect: Building a custom Redshift labeling application introduces significant operational overhead and infrastructure maintenance. Redshift is an analytical data warehouse, not an annotation workflow tool, and requires custom development for label management and data collection.
  • E is Incorrect: Amazon MSK is a managed Kafka service used for streaming data pipelines, but it does not natively trigger Ground Truth real-time labeling workflows. Ground Truth is designed for batch labeling jobs on static datasets, not continuous real-time annotation.

Key Takeaway: For labeling streaming data, buffer events into S3 via Kinesis Firehose first, then use Ground Truth's managed batch labeling workflow for human annotation with minimal operational overhead.

Question 12 - Labeling Data for Image Segmentation with SageMaker Ground Truth

An agricultural technology company is building a computer vision model to detect crop disease from drone imagery. They have 30,000 high-resolution drone images of crop fields and need to label infected regions using semantic segmentation masks. The ML engineer must choose the appropriate labeling tool and output format for SageMaker training.

Which labeling approach should the engineer use for image segmentation? (Select ONE)

  • A) Use SageMaker Ground Truth with a semantic segmentation labeling job type, where annotators draw polygon masks around infected regions, and export the augmented manifest with segmentation masks to S3.
  • B) Use SageMaker Ground Truth with a bounding box labeling job type, since bounding boxes provide sufficient annotation for segmentation models.
  • C) Use Amazon Rekognition to automatically segment infected regions without human review.
  • D) Use AWS Glue DataBrew to apply image transformations and extract segmentation masks programmatically.
  • E) Use SageMaker Data Wrangler with a custom image processing pipeline to draw segmentation masks.
Click to reveal Answer & Explanation

Correct Answer: A

Correct Option Analysis

  • A is Correct: SageMaker Ground Truth supports semantic segmentation as a built-in labeling job type. Annotators use tools to draw polygon masks on regions of interest (infected areas). The output augmented manifest includes pixel-level segmentation masks in JSON format, which SageMaker segmentation algorithms (e.g., Mask R-CNN, U-Net) consume directly for training.

Distractor Analysis

  • B is Incorrect: Bounding boxes capture object location but do not provide pixel-level boundaries needed for semantic segmentation. A segmentation model trained on bounding boxes would learn only approximate regions, losing detailed boundaries necessary for precise disease detection.
  • C is Incorrect: Amazon Rekognition detects general objects (e.g., people, vehicles, faces) but does not support custom agricultural disease segmentation. It cannot be trained on custom labels without a labeled dataset, and it cannot generate pixel-level masks for novel domains.
  • D is Incorrect: AWS Glue DataBrew is for tabular data cleaning and transformation, not image processing or segmentation mask generation. It does not have image analysis capabilities or masking tools.
  • E is Incorrect: Data Wrangler supports tabular feature transformation and some image processing via custom code, but it is not designed for human annotation workflows or pixel-level mask creation. It does not provide the built-in segmentation annotation UI that Ground Truth offers.

Key Takeaway: For image segmentation tasks, use SageMaker Ground Truth's built-in semantic segmentation labeling job type, which natively provides polygon mask drawing tools and augmented manifest outputs compatible with SageMaker segmentation algorithms.