AIF-C01 Hard Questions — 25 Difficult Questions

Question 1

A team is deploying a generative AI model for medical report generation. They must ensure patient data privacy and comply with HIPAA. Which AWS service feature is essential for de-identifying protected health information (PHI) before sending data to a foundation model?

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

Accepted Answer

Amazon Comprehend Medical. Amazon Comprehend Medical is the correct service because it is specifically designed to extract and de-identify protected health information (PHI) from unstructured medical text using natural language processing (NLP). It can detect entities such as patient names, dates, and medical record numbers, and then redact or replace them before the data is sent to a foundation model, ensuring HIPAA compliance.

Answer

AWS CloudHSM

Answer

Amazon Macie

Answer

AWS Key Management Service (AWS KMS)

Question 2

A company uses Amazon Rekognition to detect objects in images. The model is producing a high number of false positives for a specific category. Which action should be taken to improve the model's precision for that category?

Accepted Answer

Use Amazon Rekognition Custom Labels to train a custom model with additional labeled images. Amazon Rekognition Custom Labels allows you to train a custom model using your own labeled dataset, which directly addresses the false positives for a specific category by fine-tuning the model on domain-relevant examples. This improves precision because the base Rekognition model is pre-trained on general data and may not capture the nuances of your specific category, whereas a custom model learns from your labeled images to reduce incorrect detections.

Answer

Increase the confidence threshold to 99%

Answer

Submit a request to AWS to retrain the base Rekognition model

Answer

Use Amazon SageMaker to build a new object detection model from scratch

Question 3

A developer is building a RAG application on Amazon Bedrock. They notice that the model sometimes generates answers that are not supported by the retrieved documents. To reduce this, they want to enforce that the model only uses the provided context. Which Bedrock feature should they use?

Accepted Answer

Enable 'grounding' in Bedrock Guardrails. Amazon Bedrock Guardrails includes a contextual grounding check that evaluates whether a model's response is factually supported by the source context (the retrieved documents) and whether it is relevant to the user's query. By enabling grounding in Guardrails, the developer can configure thresholds that filter or flag responses with low grounding scores, directly reducing hallucinations. This is the purpose-built Bedrock feature for enforcing that answers stay within the provided context.

Answer

Increase the number of retrieved documents

Answer

Use a larger context window model

Answer

Adjust the temperature parameter to 0

Question 4

A team is fine-tuning a foundation model using SageMaker. They want to minimize training time while keeping the model's original knowledge. Which technique is BEST suited?

Accepted Answer

Use Parameter Efficient Fine-Tuning (PEFT) such as LoRA. Parameter Efficient Fine-Tuning (PEFT) methods like LoRA (Low-Rank Adaptation) are best suited because they freeze the pre-trained model weights and inject trainable low-rank matrices into specific layers, drastically reducing the number of trainable parameters. This minimizes training time and computational cost while preserving the model's original knowledge, as only a small fraction of parameters are updated during fine-tuning.

Answer

Use distributed training across multiple GPUs

Answer

Use prompt engineering instead of fine-tuning

Answer

Full fine-tuning on the new dataset

Question 5

A data scientist observes that a gradient boosting model's performance on the validation set is significantly worse than on the training set. Which adjustment is MOST likely to reduce this gap?

Accepted Answer

Reduce the learning rate and increase the number of estimators. The gap between training and validation performance indicates overfitting. Reducing the learning rate and increasing the number of estimators is a standard regularization technique in gradient boosting: a lower learning rate shrinks the contribution of each tree, forcing the model to learn more slowly and generalize better, while more trees allow the model to still capture complex patterns without overfitting to noise.

Answer

Increase the maximum depth of trees

Answer

Increase the number of features

Answer

Increase the subsample ratio to 1.0

Question 6

A company is building a resume screening model and discovers that the training data contains only resumes from one gender, leading to biased predictions. Which type of bias does this represent, and what is the most effective mitigation strategy?

Accepted Answer

Representation bias; mitigate by collecting more diverse training data. Representation bias occurs when certain groups are underrepresented in the training data. Mitigation includes collecting more diverse data or using techniques like re-weighting or synthetic data generation.

Answer

Aggregation bias; mitigate by using a single model for all groups

Answer

Measurement bias; mitigate by using more precise measurement tools

Answer

Historical bias; mitigate by removing sensitive attributes from the model

Question 7

A data scientist is fine-tuning a large language model (LLM) using Amazon SageMaker. The training job is taking a long time and the cost is higher than expected. Which configuration change would MOST effectively reduce training time and cost while maintaining model quality?

Accepted Answer

Enable managed spot training to use EC2 Spot Instances. Managed spot training uses EC2 Spot Instances, which can reduce costs by up to 90% compared to On-Demand instances. SageMaker automatically manages checkpointing and retries for spot interruptions, so model quality is maintained while training time is not negatively impacted. This directly addresses the cost concern without sacrificing performance.

Answer

Decrease the batch size

Answer

Use a larger instance type with more vCPUs

Answer

Increase the number of hyperparameter tuning jobs

Question 8

A company is building a model to predict loan default. They have historical data with 5% default rate. The model must minimize false negatives (missed defaults) because each default costs $50,000. False positives (incorrectly flagged defaults) cost $500 in customer service time. The model currently has a recall of 0.70 and precision of 0.80. Which of the following actions would MOST likely reduce the total cost?

Accepted Answer

Increase the model's recall by lowering the classification threshold. The cost of a false negative ($50,000) is 100 times greater than a false positive ($500). Lowering the classification threshold increases recall (reduces false negatives) at the expense of precision (increases false positives). Given the extreme cost asymmetry, the net expected cost will decrease even if many more false positives occur, because each additional true positive saves $50,000 while each extra false positive costs only $500. Option D directly increases recall, which is the correct lever for this cost structure.

Answer

Increase the model's precision by raising the classification threshold

Answer

Use a different algorithm that trades off recall for precision

Answer

Add more features to the model without changing the threshold

Question 9

A company wants to use a large language model to generate code based on natural language descriptions. They need to minimize latency and control costs by running inference on their own infrastructure. Which approach is most suitable?

Accepted Answer

Use Amazon SageMaker to deploy a custom LLM. Amazon SageMaker allows you to deploy a custom large language model (LLM) on your own infrastructure, giving you full control over inference latency and cost. By using SageMaker endpoints with auto-scaling and instance selection, you can optimize for low-latency responses while avoiding per-token API charges from managed services.

Answer

Use Amazon Bedrock API

Answer

Use Amazon Comprehend

Answer

Use Amazon Lex

Question 10

A company is using a RAG system with Amazon Titan Text Express for question answering. They notice that the model frequently ignores the retrieved context and generates answers based on its pre-training knowledge, leading to incorrect responses. Which change would MOST directly address this issue?

Accepted Answer

Reduce the temperature to make the model more deterministic and adhere to the context. Reducing the temperature parameter makes the model more deterministic, which increases the likelihood that it will follow the provided context rather than relying on its pre-training knowledge. In a RAG system, lower temperature forces the model to assign higher probabilities to tokens that align with the retrieved context, directly addressing the issue of ignoring context.

Answer

Switch to a larger model with more parameters

Answer

Increase the number of retrieved documents in the context window

Answer

Decrease the top-p value to 0.1 to restrict token selection

Question 11

A data scientist needs to preprocess categorical data with high cardinality (e.g., zip code with 50,000 unique values). Which technique is most appropriate?

Accepted Answer

Target encoding. Target encoding is the most appropriate technique for high-cardinality categorical features like zip codes because it replaces each category with the mean of the target variable for that category, effectively compressing 50,000 unique values into a single numeric feature while retaining predictive signal. This avoids the dimensionality explosion of one-hot encoding and the arbitrary ordering of label or ordinal encoding, which can mislead models.

Answer

Label encoding

Answer

Ordinal encoding

Answer

One-hot encoding

Question 12

A healthcare company uses Amazon SageMaker to train a model that predicts patient readmission risk based on electronic health records (EHRs) stored in Amazon HealthLake. The training dataset contains 2 million records from the past three years, with a significant gender imbalance: 70% male and 30% female. The model achieved high overall accuracy, but further analysis using SageMaker Clarify revealed that the precision for female patients is 0.65 while for male patients it is 0.88. Additionally, the model's false positive rate for female patients is significantly higher. The company must comply with healthcare regulations that require fairness and non-discrimination. The data science team has already used SageMaker Data Wrangler for initial preprocessing and SageMaker Clarify for bias detection. They need to take immediate action to mitigate the bias before deploying to production. Which course of action should the team take?

Accepted Answer

Use SageMaker Clarify's bias mitigation feature to apply reweighing techniques and retrain the model with adjusted sample weights.. To use SageMaker Clarify's built-in bias mitigation technique (reweighing) as it directly addresses the disparity in precision and false positive rates between male and female patients by adjusting sample weights during training. SageMaker Clarify's bias mitigation feature is designed for this purpose. Option B: Generating SHAP values explains predictions but does not mitigate bias. Option C: SMOTE addresses class imbalance but not fairness in terms of group accuracy disparity; it may even worsen bias by oversampling without considering fairness. Option D: SageMaker Model Monitor is for detecting drift in production data, not for mitigating bias during training. Therefore, Option A is the best choice.

Answer

Use SageMaker Clarify to generate SHAP values and adjust the model's feature importance by removing biased features.

Answer

Use SMOTE (Synthetic Minority Oversampling Technique) to balance the training dataset before retraining.

Answer

Use SageMaker Model Monitor to detect feature drift and automatically retrain the model with updated data.

Question 13

A company is using Amazon Bedrock to generate personalized marketing emails. They notice that the model sometimes produces outputs that are off-brand or contain factual errors about their products. Which TWO prompt engineering techniques would be MOST effective to address these issues? (Choose TWO.)

Accepted Answer

Include few-shot examples in the prompt that demonstrate correct brand tone and factual accuracy. Option A is correct because few-shot examples embed concrete demonstrations of the desired brand tone and accurate product facts directly in the prompt, letting the model imitate those patterns and reducing off-brand or incorrect outputs. Option C is correct because a system prompt sets persistent instructions and context—such as brand guidelines and authoritative product details—that condition every response, which directly constrains tone and factual grounding. Chain-of-thought prompting (B) mainly improves multi-step reasoning and does not supply brand or factual grounding, so it does not address these issues. Zero-shot prompting (D) removes the very examples that would steer tone and accuracy, making it counterproductive. Lowering temperature to 0.0 (E) only reduces sampling randomness; it cannot correct a model that lacks brand or product knowledge, so it does not fix off-brand or factually wrong content.

Answer

Apply chain-of-thought prompting to encourage reasoning

Answer

Use zero-shot prompting without any examples

Answer

Decrease the temperature to 0.0 to eliminate randomness

Question 14

A company fine-tunes a foundation model using SageMaker to create a domain-specific chatbot. After deployment on Bedrock, the model shows high confidence in incorrect answers. What is the most likely cause and its solution?

Accepted Answer

The model is overfitting; apply regularization techniques during fine-tuning. When a fine-tuned model exhibits high confidence in incorrect answers, it is a classic symptom of overfitting, where the model has memorized the training data too closely and fails to generalize to new inputs. Applying regularization techniques such as dropout, weight decay, or early stopping during fine-tuning reduces overfitting by penalizing overly complex decision boundaries, forcing the model to learn more robust patterns. This is the most direct cause and solution given the scenario of a domain-specific chatbot deployed on Bedrock.

Answer

The model was not pre-trained on enough data; use a larger base model

Answer

The training data was imbalanced; collect more diverse data

Answer

The inference temperature is too low; increase it

Question 15

A data scientist is using Amazon Bedrock to generate product descriptions. They notice the output is often repetitive and lacks creativity. Which combination of parameter adjustments is MOST likely to produce more diverse and less repetitive output?

Accepted Answer

Increase temperature and increase top-p. Increasing temperature raises the probability of sampling lower-probability tokens, which increases randomness and diversity. Increasing top-p (nucleus sampling) expands the set of tokens considered for sampling, further reducing repetitiveness. Together, these adjustments encourage the model to explore a wider range of possible continuations, producing more creative and less repetitive output.

Answer

Decrease temperature and decrease top-p

Answer

Decrease temperature and increase top-p

Answer

Increase temperature and decrease top-p

Question 16

Refer to the exhibit. A developer receives an error when trying to invoke the Claude Instant model from an application. The application uses the IAM role 'MyAppRole'. Which IAM policy statement should be added to the role to resolve the error?

Accepted Answer

{"Effect":"Allow","Action":"bedrock:InvokeModel","Resource":"arn:aws:bedrock:us-east-1::foundation-model/anthropic.claude-instant-v1"}. The error occurs when the application tries to invoke the Claude Instant model via the Bedrock InvokeModel API, and the IAM role 'MyAppRole' lacks the necessary permission. The required action is 'bedrock:InvokeModel', and the resource must be scoped to the specific foundation model ARN, which for Claude Instant in us-east-1 is 'arn:aws:bedrock:us-east-1::foundation-model/anthropic.claude-instant-v1'. This grants least-privilege access to invoke exactly that model.

Answer

{"Effect":"Allow","Action":"bedrock:GetFoundationModel","Resource":"*"}

Answer

{"Effect":"Allow","Action":"bedrock:InvokeModel","Resource":"arn:aws:bedrock:us-east-1::foundation-model/*"}

Answer

{"Effect":"Allow","Action":"bedrock:*","Resource":"*"}

Question 17

Which TWO practices help ensure responsible AI when deploying generative AI applications? (Select TWO.)

Accepted Answer

Implement guardrails to filter harmful or inappropriate content. Option D is correct because implementing guardrails—such as content filters, prompt shields, and safety classifiers—directly mitigates the risk of generative AI producing harmful, offensive, or inappropriate content, which is a core requirement of responsible AI deployment. Option E is correct because continuously monitoring model outputs for bias and drift ensures that the system remains fair and accurate as data distributions and user behavior change over time, enabling timely remediation. Option A is incorrect because removing content filters increases the risk of harmful outputs and violates responsible AI principles rather than ensuring them. Option B is incorrect because increasing model size at the cost of interpretability reduces transparency and explainability, which are key responsible AI goals. Option C is incorrect because using only synthetic data does not by itself guarantee privacy or fairness and can introduce its own biases and quality issues.

Answer

Deploy the model without any content filters to maximize creativity

Answer

Increase model size to improve accuracy at the expense of interpretability

Answer

Use only synthetic data for training to avoid privacy issues

Question 18

A financial services company is deploying a generative AI application using Amazon Bedrock. They need to ensure that the model does not generate responses containing personally identifiable information (PII) such as credit card numbers or Social Security numbers. The company also wants to block certain topics like investment advice. Which feature should they configure?

Accepted Answer

Amazon Bedrock Guardrails. Amazon Bedrock Guardrails is the correct feature because it provides configurable safeguards to filter and block undesirable content in foundation model responses, including PII (e.g., credit card numbers, Social Security numbers) and specific topics like investment advice. It operates at the application layer within Bedrock, allowing you to define denied topics and sensitive information filters that are enforced during inference, without requiring separate infrastructure or manual post-processing.

Answer

AWS WAF

Answer

Amazon Macie

Answer

AWS Identity and Access Management (IAM) policies

Question 19

A company is using Amazon SageMaker to manage the lifecycle of their machine learning models. They need to implement a governance framework that includes model versioning, monitoring for drift, and decommissioning of outdated models. Which THREE AWS services or features should they use together to meet these requirements? (Select THREE.)

Accepted Answer

SageMaker Pipelines. SageMaker Model Registry (C) is the correct service for model versioning and governance, as it maintains a catalog of model versions with metadata, approval statuses, and lineage, enabling tracking and controlled promotion of models through their lifecycle. SageMaker Model Monitor (E) is correct because it continuously monitors deployed endpoints for data drift, model quality drift, bias drift, and feature attribution drift, alerting when the model deviates from baseline behavior. SageMaker Pipelines (B) is correct because it provides CI/CD orchestration to automate the ML workflow, including training, evaluation, registration to the Model Registry, and deployment, which supports governance and decommissioning through repeatable, versioned pipeline executions. AWS CloudTrail (A) only records API activity for auditing and does not provide model versioning, drift monitoring, or lifecycle decommissioning, so it does not meet the requirements. SageMaker Role Manager (D) merely helps create and manage IAM roles and permissions for SageMaker personas; it does not provide versioning, drift detection, or model decommissioning capabilities.

Answer

AWS CloudTrail

Answer

SageMaker Role Manager

Question 20

A company uses Amazon Bedrock to generate product descriptions. They notice that the model sometimes produces factually incorrect information. They want to ensure responses are grounded in company-provided documents. Which Bedrock feature should they enable?

Accepted Answer

Grounding check. Bedrock Guardrails includes a grounding check that verifies the model's response is supported by the source documents (retrieved from a knowledge base). If the response is not grounded, the guardrail can block it or add a citation. The other options do not enforce factual grounding.

Answer

Topic denial

Answer

Content filtering

Answer

PII detection