AWS Certified AI Practitioner AIF-C01 (AIF-C01) — Questions 901–975

SageMaker Clarify is designed to detect bias in ML models both before and after deployment. It can analyze predictions to identify potential bias against certain groups.

Amazon Transcribe converts audio to text in real time or batch. Amazon Comprehend can then perform NLP on the transcribed text for analysis and entity extraction. Together they enable search and insights from call recordings.

Option C is correct because responsible AI guidelines, including those from AWS for Amazon Lex, require that when a chatbot cannot answer a question, it should clearly state its inability and offer alternatives (e.g., rephrasing the query or providing related topics). This maintains transparency and user trust, aligning with the 'Explainability' principle under the Guidelines for Responsible AI.

Bedrock resource policies allow you to attach fine-grained permissions to a specific model, specifying which principals (IAM roles) can invoke it.

RAG (Retrieval-Augmented Generation) allows the LLM to retrieve relevant document sections at inference time, so knowledge stays current without retraining. The other options either require expensive retraining for each update or lack document grounding.

Amazon Augmented AI (A2I) is the service designed to set up human review workflows for AI predictions, enabling human oversight when needed.

Amazon Bedrock Guardrails (Option C) directly addresses the need to prevent sensitive data from being stored in prompts by allowing you to define content filters, deny topics, and configure sensitive information filters that block or mask such data before it reaches the model. This is the native service feature designed for content safety and data leakage prevention.

The policy condition requires that the request originates from a specific VPC endpoint (vpce-12345678). If the EC2 instance is not using that endpoint (e.g., it is using a different endpoint or direct VPC access), the request will be denied. Option B is correct because the instance likely does not use that specific VPC endpoint.

Amazon Lex provides the conversational interface and natural language understanding (NLU) to interpret user intents and manage dialog, while Amazon Bedrock gives access to foundation models (FMs) for advanced natural language generation and understanding. Together, Lex can route utterances to a Bedrock FM via a Lambda function or direct integration, enabling a chatbot that leverages a pre-trained FM for richer responses.

Equalized odds requires that the model's true positive rate and false positive rate are equal across groups, which is appropriate for binary outcomes like turnover prediction.

Self-attention computes attention scores between all pairs of positions in the input, enabling the model to focus on relevant context. Feed-forward layers process each position independently; positional encoding adds sequence order information; layer normalization stabilizes training.

Dropout regularization randomly drops a fraction of neurons during training, which prevents the model from relying too heavily on specific features and forces it to learn more robust representations. This directly addresses overfitting, which is the likely cause of the loss plateauing after a few epochs in a Horovod distributed training setup on SageMaker.

Fine-tuning a pre-trained LLM on the medical dataset is cost-effective (small dataset) and adapts the model to the domain.

Document ingestion involves chunking documents into manageable pieces, generating embeddings for each chunk, and storing those embeddings in a vector store for similarity search. Prompt augmentation is part of the retrieval step, not ingestion.

The context window defines the maximum number of tokens the model can handle in one pass. To process a 6,000-token document without chunking, the model's context window must be at least 6,000 tokens (plus output).

Amazon Comprehend can perform custom classification. SageMaker allows building custom ML models. Amazon Lex can process text input and integrate with a bot to route tickets.

When classes are imbalanced, accuracy can be misleading because a model that always predicts the majority class will still have high accuracy. AUC-ROC summarizes the trade-off between true positive rate and false positive rate across thresholds, making it a robust metric for imbalanced classification. F1 score would also be good, but AUC-ROC is more commonly recommended first.

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.

Constrained generation with a controlled vocabulary is achieved by instructing the model to output only specific tokens. Few-shot examples help but do not guarantee strict adherence; output parsing via Guardrails can enforce format but not specific vocabulary.

Stability AI offers diffusion models (e.g., Stable Diffusion) for image generation from text prompts.

Model versioning tracks changes, monitoring for drift ensures performance, and a decommissioning process retires outdated models. These are key lifecycle management practices.

The pipeline should first transcribe speech to text (e.g., with Amazon Transcribe), then feed the text to an LLM for summarization.

Common automated metrics for code generation include BLEU (n-gram overlap), ROUGE (recall-oriented), and BERTScore (contextual similarity). Perplexity is for language modeling, and human evaluation is not automated. Note: While ROUGE and BLEU originate from NLP, they are also used for code generation evaluation.

Amazon SageMaker is the correct choice because it provides a fully managed environment for Jupyter notebooks and includes built-in, pre-built algorithms for binary classification. This allows the startup to quickly prototype without deep ML expertise, as SageMaker handles infrastructure, scaling, and model training.

Disparate impact measures the ratio of favorable outcomes between groups, commonly used to assess whether a selection process disproportionately affects a protected group.

RAG (Retrieval-Augmented Generation) allows the LLM to retrieve relevant document sections at inference time, so knowledge stays current without retraining. The other options either require expensive retraining for each update or lack document grounding.

The Converse API manages conversation history and context automatically, making it ideal for multi-turn chat.

Ensuring the model is invoked in the required region and data is not transferred out is critical. Bedrock does not offer customer-managed keys for model inference, and model customization may not be needed.

Custom Model Import allows you to bring your own model and fine-tune it using your own data within a VPC, keeping data fully contained.

Amazon Comprehend provides sentiment analysis and key phrase extraction; Amazon Translate can translate feedback into a common language before analysis.

In highly imbalanced datasets (99% legitimate, 1% fraud), accuracy is misleading because a model that predicts all transactions as legitimate would achieve 99% accuracy but fail to detect any fraud. The F1-score is the harmonic mean of precision and recall, providing a balanced measure that accounts for both false positives and false negatives, making it the most appropriate metric for evaluating fraud detection models.

Option A is correct because Amazon Lex endpoints support HTTPS, which encrypts data in transit using TLS. By ensuring all API calls use HTTPS, the company protects sensitive credit card information from being intercepted during communication between the client and the Lex service.

The model is overfitting, as indicated by high training accuracy but poor validation performance. Reducing training epochs or adding regularization (e.g., L1/L2 weight decay) directly addresses overfitting by limiting the model's capacity to memorize noise. In Amazon SageMaker, this can be implemented via hyperparameter tuning or by modifying the training script to include regularization terms.

A VPC-only endpoint (interface endpoint) or private endpoint without public internet access ensures traffic stays within the VPC.

Amazon Bedrock Knowledge Bases support several vector stores, including Amazon OpenSearch Serverless, Aurora pgvector, Pinecone, and MongoDB Atlas. OpenSearch Serverless is the AWS-native managed solution frequently used for this purpose.

AWS Key Management Service (KMS) allows customers to create and manage encryption keys used to encrypt data at rest in SageMaker, including training data and model artifacts.

Option A is correct because SageMaker Clarify's preprocessing transforms include a reweighting method that assigns weights to instances to adjust for fairness. Post-training (B, C) are not preprocessing. Monitoring (D) is post-deployment.

The Bedrock Converse API accepts an optional 'system' list for system prompts and a 'messages' array for conversation history, with each message having a role (user/assistant) and content. The other options misuse roles or place system prompts incorrectly.

Option B is correct because it directly addresses the data residency requirement by using a Lambda function to copy images to an S3 bucket in the desired region and then invoking Rekognition in that same region. This ensures that image data never leaves the specified AWS region, as Rekognition processes images only in the region where the API call is made. The other options either do not prevent cross-region processing or are not applicable to Rekognition's data handling.

Option C is correct because it directly addresses both issues: modifying the training container to avoid logging PHI ensures CloudWatch Logs no longer contain protected health information, and configuring the output S3 bucket to use SSE-KMS encryption ensures model artifacts are encrypted at rest. SSE-KMS is preferred over SSE-S3 for HIPAA compliance because it provides additional controls like key rotation and audit trails. This approach fixes the problems without delaying deployment, as it requires only code and configuration changes.

Choosing an embedding model that produces higher-quality vector representations improves the semantic similarity search, directly addressing retrieval relevance. The other options do not focus on retrieval quality.

Option C is correct because Amazon SageMaker Clarify is specifically designed to detect bias in machine learning models and their outputs, while Amazon Augmented AI (A2I) enables a human-in-the-loop workflow to review flagged responses. This combination directly addresses the requirement to monitor and mitigate biased or harmful responses from a generative AI chatbot, ensuring responsible AI practices.

Using provisioned throughput for predictable workloads reduces per-request cost. Choosing a smaller model variant requires less compute. Caching responses is not directly supported, and increasing max tokens increases cost.

Option A is correct because handling missing values is a critical data preprocessing step. Missing data can introduce bias or cause algorithms to fail. Imputation (e.g., using mean, median, or model-based methods) or removal of rows/columns with missing values ensures the dataset is complete and suitable for training, preventing errors during model fitting.

Cleaning text (removing punctuation, stop words), tokenization, and converting text to numerical features (e.g., TF-IDF or word embeddings) are standard preprocessing steps for text classification. Normalizing numerical features is not relevant for raw text. Encoding categorical variables may be needed if there are categorical columns, but the primary focus for text data is the textual feature engineering.

The error 'AccessDeniedException' when calling ListFoundationModels indicates that the IAM role or user executing the AWS CLI command lacks the required permission to list foundation models in Amazon Bedrock. The specific permission needed is bedrock:ListFoundationModels, which must be attached to the IAM identity via a policy. Without this permission, the API call is denied regardless of other factors like region or CLI version.

RAG (Retrieval-Augmented Generation) allows the LLM to retrieve relevant document sections at inference time, so knowledge stays current without retraining. The other options either require expensive retraining for each update or lack document grounding.

The transformer’s self-attention mechanism processes all tokens in parallel, enabling much faster training and better handling of long-range dependencies compared to sequential RNNs.

For classification tasks, accuracy (or F1 score) is the appropriate metric. ROUGE, BLEU, and BERTScore are for text generation tasks.

Word filters in Bedrock Guardrails allow you to specify custom words or phrases that should be blocked in model responses.

Amazon Comprehend is a natural language processing (NLP) service that uses machine learning to extract insights such as entities, key phrases, sentiment, and syntax from text. Amazon Lex provides NLP capabilities through automatic speech recognition (ASR) and natural language understanding (NLU) to build conversational interfaces like chatbots.

With high class imbalance, accuracy is misleading. Recall (true positive rate) measures the proportion of actual fraud cases caught, which is the key concern here.

Option C is correct because responsible AI frameworks, such as those outlined by AWS, mandate human-in-the-loop (HITL) oversight for high-stakes decisions like employee performance predictions. This ensures that automated outputs are reviewed for context, fairness, and potential errors before affecting employment outcomes, aligning with the AIF-C01 domain's emphasis on human accountability.

The Converse API expects messages as a list of previous turns. By including the full conversation history in the messages parameter, the model can reference past exchanges.

To ensure data residency, the Bedrock model must be accessed only via a VPC endpoint (so traffic stays within AWS network), and all storage and logging must be in the desired Region. CloudTrail and invocation logs must also be stored in that Region.

Option B is correct because Parameter-Efficient Fine-Tuning (PEFT) techniques like LoRA (Low-Rank Adaptation) freeze the pre-trained model weights and inject trainable rank decomposition matrices into specific layers. This drastically reduces the number of trainable parameters (often by 10,000x), lowering memory and compute requirements while preserving model quality, making it ideal for cost- and time-sensitive fine-tuning.

System prompts set behavior, content filters block harmful output, and human review catches edge cases.

One-hot encoding creates binary columns for each category, which is appropriate for linear models because it does not imply ordinal relationships. Label encoding would introduce false ordinal relationships, and feature scaling or PCA are not encoding methods.

Amazon SageMaker built-in algorithms support early stopping, which allows you to automatically terminate a training job when a specified metric, such as loss, stops improving for a defined number of consecutive epochs. This feature is configured directly in the algorithm's hyperparameters (e.g., `early_stopping_patience` for the XGBoost algorithm) and helps save compute time and cost by preventing overfitting.

Option D is correct because it combines two critical security controls: deploying the SageMaker endpoint within a VPC using an interface VPC endpoint (AWS PrivateLink) ensures that all traffic stays within the AWS network and never traverses the public internet, while configuring the model to require HTTPS enforces encryption in transit. This architecture eliminates exposure to the internet and meets the requirement for encrypted inference requests and VPC-only access.

To share a SageMaker notebook cross-account, you need to share the notebook instance or the notebook file via S3. Using a resource-based policy on the notebook instance allows cross-account access, and IAM policies can restrict actions (e.g., deny DeleteNotebookInstance). Alternatively, sharing via S3 with appropriate bucket policies and IAM roles is also valid, but the simplest approach is to use a resource-based policy on the notebook instance.

For cross-account access, the bucket policy must grant the required permissions to the SageMaker execution role (from the other account), and the SageMaker execution role must have an IAM policy that allows it to access the S3 bucket. Macie doesn't control access. VPC endpoints are optional.

RAG (Retrieval-Augmented Generation) allows the LLM to retrieve relevant document sections at inference time, so knowledge stays current without retraining. The other options either require expensive retraining for each update or lack document grounding.

An action group to query the database, a Lambda function to execute the query, and a knowledge base are essential. IAM roles are needed but not listed here. The other options are not required: fine-tuning is not needed, and custom models are optional.

Standardization (or normalization) scales features to a similar range, which is critical for linear models to avoid one feature dominating the loss due to its magnitude.

Amazon Bedrock provides a managed API to access foundation models from providers like AI21 Labs, Anthropic, and Amazon. Amazon Rekognition is for images; Textract for document text; Comprehend for natural language processing (but not foundation models per se).