View Full Amazon AWS Certified AI Practitioner AIF-C01 Exam Dumps and Practice Test Dumps.

 

Question 281

What is tokenization?

1. Splitting text into smaller units that a model can process
2. Encrypting model parameters
3. Removing all training data
4. Converting images into audio

Correct Answer: 1

Explanation

Tokenization is the process of breaking input text into smaller units called tokens so that a language model can process the text. A token may represent a whole word, part of a word, punctuation, or another text unit depending on the tokenizer. Tokenization affects how much text fits within a model’s context limit and can also influence usage-based costs when pricing depends on tokens. Because token boundaries do not always correspond directly to words, applications should not assume that a fixed number of words always produces the same number of tokens.

Question 282

A dataset contains several categorical fields with missing entries. Which approach may be appropriate before using the data for machine learning?

1. Increase model temperature
2. Apply an appropriate missing-value treatment
3. Increase network bandwidth
4. Change the model’s output language

Correct Answer: 2

Explanation

Missing categorical values should be handled according to the meaning and requirements of the dataset. Possible approaches include assigning a specific category such as “Unknown,” using an appropriate imputation method, or removing records when justified. The chosen method should avoid introducing misleading assumptions into the data. Teams should also ensure that the same preprocessing logic is applied consistently when appropriate. Missing-value handling should be evaluated using representative data because the best approach depends on why values are missing, how frequently they are missing, and how the affected feature is used by the model.

Question 283

Which technique can help identify unusually extreme numerical observations?

1. Tokenization
2. Translation
3. Outlier detection
4. Speech synthesis

Correct Answer: 3

Explanation

Outlier detection identifies observations that differ substantially from the expected pattern of a dataset. Extreme values can result from legitimate rare events, measurement errors, data-entry mistakes, or system problems. Organizations should investigate outliers rather than automatically deleting them because unusual values may contain important information. Depending on the use case, statistical methods, distance-based approaches, or domain-specific rules can be used to identify potential outliers. After detection, teams can decide whether to retain, correct, transform, or remove the observations based on their cause and relevance to the intended AI workload.

Question 284

A developer changes a model’s prompt template and wants to ensure that the new version does not accidentally expose internal instructions. What should be included in testing?

1. Only average response length
2. Only network latency
3. Only storage utilization
4. Prompt leakage testing

Correct Answer: 4

Explanation

Prompt leakage testing checks whether an AI application can unintentionally reveal internal instructions, hidden prompts, or other information that should remain protected. Testing should include attempts to manipulate the application into disclosing restricted instructions through direct and indirect requests. Developers should also review how system instructions, retrieved information, user input, and tool outputs are separated. Prompt leakage testing is only one part of application security because authorization, data protection, and output validation are still required. Security testing should be repeated whenever important prompts, models, tools, or application workflows are changed.

Question 285

What is feature selection?

1. Choosing the most relevant input features for a model
2. Generating random passwords
3. Encrypting an AI model
4. Translating training documents

Correct Answer: 1

Explanation

Feature selection is the process of identifying which available input features are most useful for a machine learning task. Removing irrelevant or redundant features can simplify a model, reduce computational requirements, and sometimes improve generalization. Feature selection differs from feature engineering, which involves creating or transforming features to represent useful information. Selection should be performed carefully to avoid using information that would not be available when the model makes real-world predictions. Evaluation should confirm that selected features contribute useful information without introducing data leakage or unnecessary complexity.

Question 286

An AI team needs humans to review selected model outputs before those outputs are used in a business process. Which AWS service is specifically designed to support human review workflows for certain AI/ML applications?

1. Amazon Route 53
2. Amazon Augmented AI (Amazon A2I)
3. Amazon EBS
4. Amazon VPC

Correct Answer: 2

Explanation

Amazon Augmented AI, also known as Amazon A2I, provides capabilities for incorporating human review into machine learning workflows. Organizations can use human reviewers to examine selected predictions or outputs when automated results require additional verification. Human review can be particularly useful when confidence is low or when decisions have important consequences. A2I can help integrate review workflows with supported AWS AI services and machine learning applications. Organizations should define clear review criteria and escalation rules so that human reviewers understand when and how outputs should be examined.

Question 287

Which activity assigns meaningful labels to examples so they can be used for supervised machine learning?

1. Data labeling
2. Model compression
3. Prompt caching
4. Network routing

Correct Answer: 1

Explanation

Data labeling assigns meaningful target information to examples so that supervised machine learning algorithms can learn from them. For example, customer messages might be labeled according to categories such as billing, technical support, or account assistance. Label quality is important because incorrect or inconsistent labels can teach the model incorrect relationships. Organizations should establish clear labeling guidelines, use appropriate quality checks, and review ambiguous examples. The amount and diversity of labeled data should also be considered because a small or unrepresentative labeled dataset may not adequately represent the conditions encountered after deployment.

Question 288

A company wants to keep an older model available while testing a newly released model version. What is the main benefit of maintaining model versions?

1. It guarantees zero downtime
2. It eliminates evaluation requirements
3. It supports controlled comparison and rollback
4. It removes all model risks

Correct Answer: 3

Explanation

Maintaining model versions allows organizations to identify exactly which model is being tested or used and can support controlled comparisons between versions. Keeping an earlier version available can also provide a rollback option if a newer version causes unexpected quality, safety, or performance problems. Version management should include relevant configuration, prompts, evaluation results, and deployment information where appropriate. This creates better traceability throughout the model lifecycle. Versioning does not eliminate model risks, but it makes changes easier to manage, evaluate, document, and reverse when necessary.

Question 289

Which practice helps determine whether a model remains useful after deployment?

1. Ongoing evaluation
2. Removing all monitoring
3. Ignoring user feedback
4. Changing the model randomly

Correct Answer: 1

Explanation

Ongoing evaluation helps determine whether an AI model continues to meet established quality, safety, and performance requirements after deployment. Real-world conditions can change because user behavior, data distributions, business requirements, or external information may change over time. Organizations can monitor appropriate metrics, collect feedback, review representative examples, and periodically reevaluate the model. Monitoring should have defined thresholds and response procedures so that identified problems lead to appropriate action. Continuous evaluation is especially important for systems whose outputs influence important business processes or user-facing decisions.

Question 290

A generative AI application receives a large number of requests containing irrelevant text that increases processing costs. What should developers consider first?

1. Increasing irrelevant context
2. Expanding every prompt
3. Removing useful instructions
4. Filtering unnecessary input content

Correct Answer: 4

Explanation

Filtering unnecessary input content can reduce the amount of irrelevant information sent to a model. This may improve efficiency, reduce token usage, and sometimes improve response quality by keeping the model’s context focused on information relevant to the task. Developers should avoid removing information that is necessary for accurate processing. Input filtering can include removing duplicate content, irrelevant fields, or unnecessary historical information according to defined application rules. Teams should measure the effect of filtering because excessive reduction of context can also negatively affect the quality or completeness of generated responses.

Question 291

Which AWS service can detect sensitive information such as personally identifiable information in data stored in Amazon S3?

1. Amazon Macie
2. Amazon Polly
3. Amazon Lex
4. Amazon CloudFront

Correct Answer: 1

Explanation

Amazon Macie is designed to help organizations discover and protect sensitive data stored in Amazon S3. It can identify potentially sensitive information, including certain types of personally identifiable information, and can provide visibility into the security and privacy characteristics of S3 data. This capability can help organizations identify data that requires additional controls or review. Macie is not a replacement for access controls, encryption, or data-governance policies. Instead, it provides discovery and analysis capabilities that can support broader data-protection programs.

Question 292

What does data provenance describe?

1. The speed of a model response
2. The origin and history of data
3. The number of model parameters
4. The size of an image

Correct Answer: 2

Explanation

Data provenance describes where data came from and, depending on the implementation, how it was collected, transformed, processed, or transferred over time. Maintaining provenance can help organizations understand the sources behind training and evaluation datasets and investigate problems when unexpected results occur. Provenance information can also support governance, reproducibility, auditing, and data-quality investigations. Organizations should document meaningful transformations and source information throughout the data lifecycle. Strong provenance practices are particularly valuable when datasets are assembled from multiple systems or are repeatedly transformed before being used by AI applications.

Question 293

A model is tested only with clean, ideal inputs, but real users may provide misspelled, incomplete, or unusual requests. What type of testing would better reflect this situation?

1. Storage testing
2. DNS testing
3. Robustness testing
4. Billing testing

Correct Answer: 3

Explanation

Robustness testing evaluates how well a model performs when inputs differ from ideal or expected examples. For AI applications, this can include misspellings, incomplete requests, unusual phrasing, noisy information, unexpected formats, or other realistic variations. Testing these conditions can reveal weaknesses that are hidden when evaluation uses only clean examples. The test set should reflect the actual application environment and should include cases that are reasonably expected after deployment. Robustness testing can help organizations decide whether additional preprocessing, prompt changes, model changes, or fallback mechanisms are necessary.

Question 294

A company wants to control how much an AI API can be called by a particular application over a given period. Which mechanism is most relevant?

1. Rate limiting
2. Image resizing
3. Model labeling
4. Data anonymization

Correct Answer: 1

Explanation

Rate limiting controls the number or frequency of requests an application can make during a specified period. It can help protect AI services from excessive request volume, manage costs, reduce accidental overload, and support fair resource usage. Rate limits can be applied at different levels depending on the architecture, such as per user, application, API key, or service. Applications should handle rejected or throttled requests appropriately rather than continuously retrying without limits. Rate limiting works alongside authentication, authorization, monitoring, and service-specific quotas as part of a broader API protection strategy.

Question 295

Why might an organization use a smaller model instead of a larger foundation model?

1. To guarantee perfect responses
2. To remove the need for testing
3. To potentially reduce cost and resource requirements
4. To eliminate all security controls

Correct Answer: 3

Explanation

A smaller model may require fewer computational resources and can potentially provide lower latency and lower inference costs when it meets the application’s quality requirements. Model selection should be based on measured performance rather than model size alone. A larger model may provide better results for some complex tasks, while a smaller model may be sufficient for simpler workloads. Organizations should compare candidate models using representative inputs and criteria such as quality, safety, latency, and cost. Choosing an appropriately sized model can support efficient AI architecture without unnecessarily increasing operational expenses.

Question 296

An organization wants to ensure that AI-generated content is reviewed before being used for a high-impact decision. Which approach is appropriate?

1. Fully automate the decision without controls
2. Require human oversight before the consequential action
3. Remove all evaluation data
4. Increase generation randomness

Correct Answer: 2

Explanation

Human oversight can provide an additional safeguard when AI outputs are used in high-impact or consequential decisions. A qualified reviewer can assess whether the generated information is appropriate and whether additional evidence or investigation is needed before an action is taken. Organizations should define when human review is mandatory, what reviewers should examine, and how disagreements or uncertain cases are handled. Human oversight does not automatically make an AI system safe or accurate, so it should operate alongside testing, authorization, documentation, monitoring, and other responsible AI controls appropriate to the use case.

Question 297

Which concept describes the ability of a model to perform well on new examples that were not part of its training data?

1. Generalization
2. Tokenization
3. Encryption
4. Quantization

Correct Answer: 1

Explanation

Generalization is the ability of a machine learning model to perform effectively on new, unseen examples rather than only memorizing its training data. Good generalization is important because real-world inputs are typically different from the exact examples used during training. Organizations can assess generalization using separate validation or test data that was not used to train the model. Representative evaluation data is important because a model may generalize well to one population or scenario but perform poorly under different conditions. Overfitting is a common problem that can reduce generalization.

Question 298

A team wants to reduce the amount of sensitive information contained in a dataset before sharing it with an internal analytics group. Which action is appropriate?

1. Add more personal identifiers
2. Remove or mask unnecessary sensitive information
3. Increase the model’s context size
4. Duplicate the dataset

Correct Answer: 2

Explanation

Removing or masking unnecessary sensitive information can reduce privacy exposure before data is shared. The organization should first determine which information is required for the intended analytics purpose and then apply appropriate protection methods to information that is not necessary. Depending on the situation, this may involve removing direct identifiers, masking values, aggregating information, or applying other privacy-preserving techniques. Access controls should remain in place because transformed data may still contain sensitive information. Data protection should be designed according to the sensitivity, purpose, and applicable requirements of the dataset.

Question 299

Which AWS capability can help create alerts when monitored AI application metrics exceed defined thresholds?

1. Amazon CloudWatch alarms
2. Amazon Polly speech marks
3. Amazon Textract queries
4. Amazon Translate terminology

Correct Answer: 1

Explanation

Amazon CloudWatch alarms can evaluate monitored metrics against defined thresholds and initiate configured actions or notifications when conditions are met. For AI applications, teams can use monitoring and alarms to identify issues such as unusually high latency, resource utilization, request failures, or other operational conditions. Effective alarms should have meaningful thresholds to avoid excessive false alerts. Organizations should also define what action should follow an alert, such as investigation, scaling, rollback, or escalation. CloudWatch alarms complement broader application monitoring and should be designed around the operational requirements of the AI workload.

Question 300

A team wants to verify that an AI system continues to meet its safety requirements after a model update. What should it do?

1. Skip testing because the application already worked
2. Test only the user interface
3. Re-run relevant safety evaluations against the updated system
4. Delete previous evaluation results

Correct Answer: 3

Explanation

Relevant safety evaluations should be repeated after a significant model update because changes in model behavior can affect safety characteristics even when the surrounding application remains unchanged. Testing should use representative and challenging scenarios that reflect the application’s safety requirements. Previous evaluation results should be retained so that the team can compare the updated model with earlier versions and identify meaningful changes. Depending on the use case, organizations may also perform human review, adversarial testing, and regression testing. Safety validation should be treated as an ongoing lifecycle activity rather than a one-time deployment step.