AIF-C01 Practice Questions — Free

Question 1 of 25

What is the difference between AI and ML?

• A.They are unrelated fields with no overlap in methods
• B.AI refers only to robotics, while ML refers only to statistics
• C.ML is a subset of AI focused on learning patterns from dataCorrect
• D.AI is a subset of ML focused on rule-based decision logic

Question 2 of 25 · multiple response

Which of the following are examples of unsupervised learning? (Select TWO)

• A.Topic discovery in documentsCorrect
• B.Spam classification with labeled emails
• C.House price regression
• D.Customer segmentation via clusteringCorrect

Question 3 of 25

A shipping company wants to estimate delivery time in hours for each package. What model output type is this?

• A.A cluster identifier from an unsupervised model
• B.A continuous value from a regression modelCorrect
• C.A reward signal from a reinforcement learning agent
• D.A discrete class label from a classification model

Question 4 of 25

A team must convert raw customer records into numeric inputs suitable for a model. What is this stage of the ML lifecycle?

• A.Model deployment to an endpoint
• B.Model monitoring for drift
• C.Real-time inference on new data
• D.Feature engineeringCorrect

Question 5 of 25

A team wants the model to learn from streaming data continuously rather than retraining from scratch. What concept is this?

• A.Transfer learning from another domain
• B.Online or incremental learningCorrect
• C.Batch training on a fixed dataset
• D.Zero-shot inference without training

Question 6 of 25

A model trained on data up to a certain date cannot answer about later events. What is this limitation called?

• A.Context window limit
• B.Knowledge cutoffCorrect
• C.Hallucination tendency
• D.Inference latency

Question 7 of 25

A prompt is written to ask the model to critique and revise its own first answer. What technique is this?

• A.Zero-shot prompting without examples
• B.Self-refinement or self-critique promptingCorrect
• C.Data augmentation of the training set
• D.Model distillation into a smaller model

Question 8 of 25

A team wants a model to translate documents while preserving formatting like tables. Which model capability matters?

• A.Speech synthesis of the translated document text
• B.Image generation of the translated document pages
• C.Time-series forecasting of translation demand
• D.Instruction following with structured output supportCorrect

Question 9 of 25

A team fine-tunes a model and then finds it performs worse on general tasks it previously handled well. What is this called?

• A.Training data leakage
• B.Hallucination in outputs
• C.Catastrophic forgettingCorrect
• D.Prompt injection by users

Question 10 of 25

A company wants to understand what 'fine-tuning' changes in a model. What is correct?

• A.It adjusts the model's weights using task-specific dataCorrect
• B.It increases the size of the model's context window
• C.It changes only the prompt template used at runtime
• D.It changes the API endpoint the model is served from

Question 11 of 25

A team notices an LLM's knowledge stops at a certain date. What is this limitation called?

• A.Temperature ceiling
• B.Token budget exhaustion
• C.Training data cutoffCorrect
• D.Context window limit

Question 12 of 25

A team wants to constrain an agent to complete tasks within a fixed number of reasoning steps. What does this control?

• A.The IAM policy attached to the agent's execution role
• B.Cost and runaway loops during agent orchestrationCorrect
• C.The parameter count of the agent's foundation model
• D.The size of the vector database the agent queries

Question 13 of 25

A RAG application returns irrelevant documents. What component most likely needs improvement?

• A.The output token limit on the model's responses
• B.The temperature setting on the generation step
• C.The retrieval step, such as chunking or embedding qualityCorrect
• D.The IAM role attached to the application

Question 14 of 25

A team wants to import a model they fine-tuned elsewhere into Bedrock for inference. What capability supports this?

• A.Model Evaluation jobs
• B.Custom Model ImportCorrect
• C.Guardrails for content filtering
• D.Knowledge Bases for retrieval

Question 15 of 25

A developer wants to use a foundation model from a Python application. What do they use?

• A.Amazon Polly's Python client library
• B.SSH access to the model's underlying host
• C.The AWS SDK to call the Bedrock InvokeModel APICorrect
• D.The AWS Console's playground interface only

Question 16 of 25

A team fine-tunes a model but has only 200 examples. What is a likely risk?

• A.Overfitting to the small dataset with poor generalizationCorrect
• B.Loss of IAM permissions on the fine-tuned model
• C.The model becoming too large to deploy on an endpoint
• D.Excessive inference cost due to the added examples

Question 17 of 25

A team wants a model to draft responses that an agent reviews before sending. What pattern is this?

• A.Zero-shot generation with no human review at all
• B.Fully autonomous responses sent without review
• C.Human-in-the-loop with AI-generated draftsCorrect
• D.Batch inference of all responses overnight only

Question 18 of 25

A team wants to extract key-value pairs from insurance claim forms at scale. Which service fits?

• A.Amazon Polly
• B.Amazon Comprehend
• C.Amazon TextractCorrect
• D.Amazon Lex

Question 19 of 25

A team wants documentation of a model's intended use, limitations, and evaluation results. What SageMaker feature supports this?

• A.SageMaker Ground Truth
• B.SageMaker Model CardsCorrect
• C.SageMaker Canvas
• D.SageMaker Pipelines

Question 20 of 25

A company wants to ensure AI-generated customer messages are reviewed for tone and accuracy. Which principle does this support?

• A.Redundancy of the message store
• B.Elasticity of the messaging system
• C.Low latency of message delivery
• D.Human oversight and quality controlCorrect

Question 21 of 25

A model used for credit decisions must not use protected attributes like race. What practice addresses this?

• A.Allocating more GPUs to the inference endpoint
• B.Raising the temperature to diversify the decisions
• C.Increasing the model's size to capture more nuance
• D.Excluding protected attributes and testing for proxy variablesCorrect

Question 22 of 25

A team notices a model's outputs are less accurate for non-native English speakers. What should they do?

• A.Ignore the difference as an acceptable model limitation
• B.Reduce logging to lower the ongoing cost of monitoring
• C.Evaluate by user group and improve data representationCorrect
• D.Increase the temperature to diversify the model's outputs

Question 23 of 25

A company wants to prevent developers from exporting fine-tuned model weights out of AWS. What control helps?

• A.IAM policies restricting model export and artifact accessCorrect
• B.A larger context window on the fine-tuned model
• C.Public S3 access to the model artifact bucket
• D.A higher temperature on the fine-tuned model

Question 24 of 25

An organization wants to prevent employees from using unapproved external AI tools with company data. What control helps?

• A.Larger context windows on the approved internal models
• B.Public S3 buckets for sharing prompts across teams
• C.Higher temperature settings on the approved internal models
• D.Data loss prevention policies and approved internal AI toolsCorrect

Question 25 of 25

A company wants only certain teams to invoke a specific Bedrock model. What enforces this?

• A.Raising the model's temperature for those teams only
• B.A public API gateway placed in front of the model
• C.A larger context window configured for the model
• D.IAM policies scoping bedrock:InvokeModel to that model's ARNCorrect