Task 1.3: Describe the AI/ML development lifecycle.

• ASageMaker Feature Store

  Incorrect

  Feature Store creates, stores and shares features, the inputs to models. It does not track model versions or approvals.

• BSageMaker Model Registry

  Correct

  Model Registry catalogs models for production, manages model versions and their metadata such as training metrics, manages approval status and deploys models to production.

• CSageMaker Data Wrangler

  Incorrect

  Data Wrangler imports, prepares, transforms and analyzes data before training. It is a data preparation tool, not a model catalog.

• DSageMaker JumpStart

  Incorrect

  JumpStart provides pretrained, open-source models and solution templates to get started. It does not govern the team's own versions.

Versioning, approval and promotion of trained models belong to the deployment stage of the lifecycle, which SageMaker Model Registry handles.

• APrecision

  Incorrect

  Precision measures how many positive predictions were right. It matters most when a false positive is costly, which is the less expensive error here.

• BAccuracy

  Incorrect

  Accuracy counts all correct predictions together. For a rare condition the data is imbalanced, and AWS points to balanced accuracy as the better measure in that case; neither targets missed positives directly.

• CRoot mean squared error (RMSE)

  Incorrect

  RMSE measures the error of numeric predictions in regression. This is a classification problem.

• DRecall

  Correct

  Recall is the share of actual positives the model finds, TP / (TP + FN). It is the metric to raise when false negatives are costly, such as in cancer testing.

Choose the metric by the cost of each error: costly false negatives call for recall, costly false positives for precision, and F1 balances the two.

• AAutomate the stages from data preparation through training, validation and deployment so each run is repeatable

  Correct

  MLOps automates the stages of the ML pipeline, from data ingestion and preprocessing through training and validation to deployment, for repeatability, consistency and scale.

• BDeploy each model once and avoid retraining it, so that its behavior in production never changes

  Incorrect

  MLOps includes continuous monitoring and continuous training: models are retrained and redeployed as data and results change.

• CVersion the training code, data and models so that results can be reproduced and rolled back

  Correct

  Version control of ML assets makes training reproducible and auditable and lets the team roll back to a previous version.

• DKeep each scientist's experiments private to that person's notebook so the results stay independent

  Incorrect

  MLOps tracks experiments and manages training pipelines centrally; private, untracked experiments are the problem it replaces.

• EPromote a model to production as soon as an experiment finishes, without recording any test results

  Incorrect

  MLOps adds automated testing and validation before release, so problems are found early and releases are auditable.

MLOps brings software delivery habits to ML: automated, repeatable pipelines, versioned assets, testing, and continuous monitoring with retraining.

• AHost an open-source model on Amazon EC2 instances that the startup patches and scales itself

  Incorrect

  On Amazon EC2 the customer manages the guest operating system, patches and installed software, which is exactly the work the startup wants to avoid.

• BDeploy a model from SageMaker JumpStart to a real-time endpoint backed by an instance type the startup chooses

  Incorrect

  This is a self-hosted endpoint: real-time inference uses a persistent endpoint backed by the instance type you choose, so there is an endpoint to size and run.

• CCall a foundation model through Amazon Bedrock, a serverless managed API

  Correct

  Amazon Bedrock is serverless, so there is no infrastructure to manage, and one API gives access to foundation models from several providers.

• DPre-train a new foundation model on the startup's own documents before using it

  Incorrect

  Foundation models are already trained on broad data and can perform general tasks such as summarization; pre-training a new one is the most expensive path and needs ML expertise.

A managed API service is the lowest-effort way to use a foundation model in production; self-hosting trades that convenience for more control.

• ASageMaker Model Registry

  Incorrect

  Model Registry catalogs model versions and their approvals. It does not store feature values.

• BSageMaker Feature Store

  Correct

  Feature Store creates, stores and shares features for training and inference and ingests them in a consistent way, which reduces training-serving skew.

• CSageMaker JumpStart

  Incorrect

  JumpStart offers pretrained models and solution templates; it does not manage the team's features.

• DAmazon SageMaker Canvas

  Incorrect

  Canvas lets users generate predictions without code. It is not a shared store of features for existing pipelines.

A feature store is the single source of truth for model inputs, so training and inference see the same values.

• ASageMaker Feature Store

  Incorrect

  Feature Store stores and shares features. It does not orchestrate the steps of the workflow.

• BSageMaker Serverless Inference

  Incorrect

  Serverless Inference hosts a trained model for predictions; it does not run training workflows.

• CSageMaker Pipelines

  Correct

  Pipelines is a purpose-built workflow orchestration service that automates ML development on serverless infrastructure that SageMaker AI provisions and scales.

• DSageMaker Model Cards

  Incorrect

  Model Cards document a model's intended use, risk and evaluation for governance; they do not run steps.

Repeatable, automated workflows are the core of MLOps, and in SageMaker AI they are built with Pipelines.

• AF1 score

  Correct

  F1 is the harmonic mean of precision and recall, so it rises only when both are good.

• BRecall

  Incorrect

  Recall covers only the first goal, finding the actual positives; on its own it can be maximized by flagging everyone.

• CPrecision

  Incorrect

  Precision covers only the second goal, the quality of positive predictions.

• DRoot mean squared error (RMSE)

  Incorrect

  RMSE measures the error of numeric predictions in regression, not the quality of class labels.

Recall asks "did we find them?", precision asks "were we right when we said so?", and F1 combines the two in one number.

• AFreeze the model so its behavior stays the same over time

  Incorrect

  A frozen model keeps making the same mistakes as the world moves on; MLOps plans for monitoring and retraining instead.

• BReplace the model with a larger one trained on the same historical data

  Incorrect

  A bigger model trained on the same old data does not learn the new behavior the forecasts are missing.

• CEvaluate the model only once a year to reduce operating cost

  Incorrect

  Less frequent evaluation would hide the problem longer; MLOps calls for continuous monitoring.

• DMonitor model quality in production and retrain on recent data

  Correct

  MLOps includes continuous monitoring with metrics and continuous training that automatically retrains models for redeployment, and monitoring events can trigger retraining.

Models decay as the world changes. Monitoring catches it, and automated retraining on fresh data fixes it.

• APrecision of the model's predictions

  Incorrect

  Precision is a model performance metric: the share of positive predictions that are correct.

• BCost per user of running the model

  Correct

  The exam guide lists cost per user among the business metrics used to evaluate ML models.

• CF1 score on the test set

  Incorrect

  F1 is a model performance metric that combines precision and recall.

• DReturn on investment of the project

  Correct

  Return on investment (ROI) is one of the business metrics the exam guide lists.

• EAccuracy on the validation set

  Incorrect

  Accuracy is a model performance metric: the ratio of correct predictions to all predictions.

Model metrics say how well the model predicts; business metrics say whether the project pays off. A good model can still be a bad investment.