AWS Certified AI Practitioner AIF-C01 (AIF-C01) — Questions 601–675

Amazon Transcribe is the correct choice because it is a fully managed automatic speech recognition (ASR) service designed specifically to convert speech into text. It can handle real-time streaming or batch processing of audio files, making it ideal for transcribing customer service calls into searchable text.

Amazon Bedrock is a fully managed service that provides a single API to access and experiment with multiple foundation models from leading AI providers like AI21 Labs, Anthropic, Cohere, Meta, Stability AI, and Amazon itself. This allows developers to quickly test different models without managing underlying infrastructure or learning separate APIs for each provider.

The 'CUDA out of memory' error indicates that the GPU memory is exhausted during training. Reducing `per_device_train_batch_size` directly decreases the number of samples processed simultaneously per GPU, which lowers memory consumption for activations, gradients, and optimizer states. This is the most direct and effective hyperparameter change to resolve an out-of-memory condition.

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 Aurora with the pgvector extension (Option D) enables storing and querying vector embeddings directly within a PostgreSQL-compatible database, which is essential for Retrieval-Augmented Generation (RAG). When combined with Amazon Bedrock Knowledge Bases (Option E), it allows the company to retrieve the most current market data as vector embeddings, ensuring the generated investment summaries are grounded in factual, up-to-date information rather than relying solely on the model's static training data.

Increasing the training data size and using regularization techniques like dropout or weight decay are standard ways to combat overfitting. Decreasing learning rate or increasing epochs can worsen overfitting; more layers usually increase overfitting.

AWS CloudTrail logs all API activity in SageMaker Studio, including user actions and access attempts, while Amazon GuardDuty provides intelligent threat detection by analyzing CloudTrail logs, VPC flow logs, and DNS logs for unauthorized access patterns. Together, they meet compliance requirements for logging and detecting unauthorized access without additional configuration overhead.

Increasing the volume of event data provides Amazon Fraud Detector with more examples of both fraudulent and legitimate transactions, which allows the model to learn more robust patterns and reduce overfitting. More data helps the model generalize better to unseen events, directly improving prediction accuracy.

Including relevant context from the knowledge base and providing clear system instructions improve accuracy. Other options do not directly help.

Option B is correct because it addresses all three requirements: SSE-KMS encrypts data at rest in S3, SageMaker inter-container traffic encryption ensures data in transit between training containers is encrypted, and CloudTrail data events for the S3 bucket log all access to the PII data. This combination provides a complete, auditable encryption and logging solution.

Amazon Kendra is a highly accurate enterprise search service that can retrieve relevant documents from various sources, which can then be provided to a foundation model for generation. Lex, Connect, and Polly are not designed for document retrieval.

Option A is correct because foundation models have a fixed maximum context window (e.g., 4,096 tokens for GPT-3.5). By breaking a long document into smaller chunks and summarizing each independently, you avoid truncation and ensure the model can process the entire content without losing information. This chunking strategy is a standard preprocessing technique for handling documents that exceed the model's context length.

Embedding models convert text into dense vector representations that capture semantic meaning, enabling similarity search in vector databases.

ROUGE (Recall-Oriented Understudy for Gisting Evaluation) measures n-gram overlap between generated and reference summaries. BLEU is for translation, BERTScore uses embeddings, and perplexity measures language model fluency.

Amazon Bedrock Model Evaluation with toxicity metrics is the most effective approach because it provides automated, built-in evaluation of model outputs for toxic content using predefined metrics, directly integrated with the Bedrock service. This eliminates the need for manual effort or custom filtering, ensuring consistent and scalable monitoring of harmful content.

Transparency requirements often mandate disclosure that content is AI-generated, which can be done through labels, watermarks, or metadata. Providing a clear disclosure to users is the most direct and compliant approach.

Amazon GuardDuty is a threat detection service that continuously monitors AWS accounts and workloads using machine learning, anomaly detection, and integrated threat intelligence. It specifically analyzes CloudTrail management and data events, VPC Flow Logs, and DNS logs to identify unauthorized behavior or potential security threats, making it the correct choice for automatically detecting anomalies in CloudTrail logs.

Grounding check ensures the model's response is grounded in the retrieved documents, penalizing or filtering content that is not supported. Content filtering obscenity, topic denial, and PII detection do not address hallucination of facts.

Standardization (Z-score scaling) transforms features to have a mean of 0 and a standard deviation of 1, making it a valid type of feature scaling. It is essential when using algorithms that assume normally distributed data, such as linear regression, SVM, or PCA, and it does not bound the data to a fixed range.

Amazon Comprehend automatically encrypts all data at rest using AES-256 encryption by default, with no additional configuration required. This encryption covers both the documents processed by the service and any models or artifacts stored internally. Therefore, no action is needed from the customer to enable encryption at rest.

Option D is correct because directly instructing the model to avoid specific security vulnerabilities (e.g., SQL injection, buffer overflows) is the most explicit and effective way to constrain the output. Amazon Bedrock models respond well to clear, imperative instructions in the system prompt or user message, making this a direct application of prompt engineering for safety. Chain-of-thought or few-shot examples may improve reasoning or style but do not guarantee the model will avoid insecure patterns unless explicitly told to do so.

Using SageMaker's managed spot training can reduce cost but not necessarily time. Incremental training is not a standard SageMaker feature. The most direct way to reduce time is to increase the number of instances for distributed training, provided the model supports data parallelism.

For unlabeled data with rare anomalies, a semi-supervised approach using a one-class classifier is appropriate because it learns the 'normal' pattern. Amazon Lookout for Equipment is designed for sensor data anomaly detection and can be used in a semi-supervised manner. The other options require labels or are not suited for real-time IoT data.

The correct answer is fairness because the chatbot's responses are biased and offensive, indicating a lack of fairness in treating all users appropriately. Accountability relates to ownership, privacy to data protection, and transparency to explainability.

The model performs exceptionally well on training data (99% accuracy) but significantly worse on validation data (70% accuracy). This large gap indicates the model has memorized the training data, including noise and irrelevant patterns, rather than learning generalizable features — a classic symptom of overfitting.

The model is overfitting: it has memorized the training data but fails to generalize. Overfitting occurs when a model is too complex relative to the amount of training data. The other options describe underfitting (not learning from training data) or data leakage (causing overly optimistic training but not necessarily poor unseen performance).

Amazon SageMaker provides real-time inference endpoints that allow you to deploy trained models to a fully managed HTTPS endpoint for low-latency predictions. These endpoints automatically scale based on traffic and support A/B testing, making them suitable for production workloads.

Option A is correct because configuring a SageMaker endpoint to use a custom SSL certificate from AWS Certificate Manager (ACM) ensures that all data transmitted between clients and the endpoint is encrypted in transit using TLS. This enforces the compliance requirement for encryption in transit by replacing the default SageMaker certificate with a customer-managed certificate, which can be validated and rotated as needed.

Amazon Bedrock Guardrails provide content filtering, allowing you to define policies to block harmful or inappropriate content. Prompt templates and temperature affect output style but not safety. IAM controls access but not content.

Option C is correct because disabling 'Direct Internet Access' on a SageMaker notebook instance and placing it in a VPC without an internet gateway or NAT device ensures the notebook cannot reach the public internet. This configuration forces all traffic to stay within the VPC, preventing data exfiltration via outbound internet connections while still allowing access to VPC resources.

Bedrock Guardrails provide content filters, topic denial, PII detection, and other safety controls. Knowledge Bases, Agents, and Playground do not directly enforce content safety rules.

Streaming responses allow the model to send tokens as they are generated, enabling the application to display incremental output to the user, which reduces perceived latency.

A system prompt sets the overall behavior, persona, and constraints for the model's responses, guiding how it should interact.

Amazon Titan Embeddings is a text embedding model that produces vector representations. Titan Text models generate text; Titan Image Generator creates images; Titan Multimodal Embeddings handles both text and images but the question asks for text-to-vectors specifically.

Amazon GuardDuty uses machine learning (including anomaly detection) to analyze CloudTrail logs, VPC flow logs, and DNS logs, and it works out-of-the-box without manual labeling. Macie is for sensitive data, Detective for investigation, and SageMaker would require custom model building.

Model caching in Amazon Bedrock stores recently generated responses for identical prompts and reuses them, reducing the number of inference calls and thus costs. Batch inference processes multiple requests together but still charges per request. Prompt management and fine-tuning do not directly reduce per-request costs.

Guardrails can filter harmful content, but for code-specific vulnerabilities, a more targeted approach is to fine-tune on secure code examples or use a specialised model. However, among the options, applying guardrails for code safety is a practical step.

Option C is correct because SageMaker's model parallelism strategy with the SageMaker distributed training library is specifically designed for training large models that do not fit into the memory of a single GPU. It partitions the model layers across multiple GPUs, enabling efficient training of models with hundreds of billions of parameters by overlapping computation and communication.

Options A and B are correct: Diverse training data and bias detection with Clarify. Options C, D, E are not direct fairness actions.

The `max_tokens` parameter directly controls the maximum number of tokens (words or subwords) the model can generate in a single response. By reducing this value, the data scientist caps the output length, preventing overly long responses. Temperature and top_p affect randomness and diversity, not length, while stop sequences define when generation halts but do not enforce a hard token limit.

D is correct because 'Invocations per instance' is a custom metric that directly measures the number of inference requests each instance is handling. By setting a target value for this metric, the scaling policy can proactively add instances when the per-instance request count approaches a threshold, preventing latency increases before they occur. This is the recommended approach for SageMaker endpoints with gradual traffic increases, as it anticipates demand rather than reacting to latency spikes.

Using a smaller instance type reduces latency, auto scaling handles variable traffic, and enabling data capture allows monitoring and debugging.

SageMaker endpoints can be deployed in a VPC and configured to use a private VPC endpoint (AWS PrivateLink) so that they are accessible only from within the VPC. HTTPS ensures encryption in transit; SageMaker endpoints automatically use HTTPS.

Model distillation reduces the size of the foundation model by training a smaller 'student' model to mimic the behavior of a larger 'teacher' model. This directly decreases inference latency because the smaller model requires fewer computational resources (FLOPs) per forward pass, which is critical for real-time translation in a chat application where low latency is paramount.

Amazon SageMaker Ground Truth supports private workforce labeling where company employees (e.g., via a corporate directory or invited users) perform manual annotation. This is ideal for sensitive data or domain-specific tasks like custom object detection, where internal expertise ensures high label accuracy.

The test set is used to evaluate the final model's generalization performance on unseen data. The validation set is for hyperparameter tuning during development. Training set is for fitting the model.

Option A is correct because increasing the sensitivity of the content filters in Bedrock Guardrails directly addresses the issue: the current filters are set to the lowest sensitivity, allowing speculative or promise-based responses to pass through. By raising the sensitivity, the guardrails will block more non-compliant content while still permitting legitimate investment advice, striking the required balance between regulatory compliance and functionality.

The error indicates missing permission to invoke the specific model. The correct action is 'bedrock:InvokeModel' on the specific model ARN.

Bedrock Playground is a no-code interface within Bedrock Studio for experimenting with models and prompts. Knowledge Bases, Agents, and Guardrails are not designed for rapid prototyping of prompts and models.

Cohere Embed supports multiple languages natively, while Titan Text Embeddings is primarily English-focused.

Option C is correct because the IAM policy includes a condition that enforces the `DataCaptureConfig.EnableCapture` parameter to be set to `true` when creating a SageMaker endpoint. This ensures that model monitoring data is automatically collected, which is a key practice for responsible AI as it allows continuous monitoring of model performance, bias detection, and drift analysis. Without data capture, teams cannot audit or validate model behavior in production, undermining accountability and transparency.

SageMaker Clarify can compute SHAP values for per-prediction explanations. SageMaker Pipelines can orchestrate the generation of reports on a schedule. Data Wrangler is for data preparation, Model Monitor for drift, Ground Truth for labeling – none generate per-prediction explanation reports automatically.

The error indicates the SageMaker execution role lacks permission to decrypt the KMS key. The most direct solution is to add the role to the KMS key policy with the kms:Decrypt permission. Option B is unnecessary; Option C changes encryption approach unnecessarily; Option D is also valid but less direct because the key policy must allow the role anyway, and modifying the key policy is the standard practice for SageMaker.

Temperature controls randomness; lower values make the model more deterministic by sharpening the probability distribution.

Amazon SageMaker Autopilot is the correct choice because it automatically performs data preprocessing (including handling missing values), feature engineering, model selection, and hyperparameter tuning for supervised learning tasks like binary classification. It requires minimal code—users can simply point to a tabular dataset in Amazon S3 and specify the target column, and Autopilot will automatically train and evaluate multiple candidate models, making it ideal for quickly building a binary classifier on a 10,000-row, 200-feature dataset with missing values.

One-class SVM is specifically designed for anomaly detection, as it learns a boundary around the normal data points in the feature space and identifies any point falling outside this boundary as an anomaly. This makes it ideal for detecting unusual patterns in server metrics without requiring labeled anomaly examples.

Amazon OpenSearch Serverless is a fully managed AWS vector database. The others are either not AWS-managed or not vector stores.

SHAP values provide per-instance (local) explanations of feature contributions. Permutation feature importance, partial dependence plots, and global feature importance all give global explanations, not local.

Decreasing the temperature (Option E) reduces randomness by lowering the probability of sampling lower-ranked tokens, making the model more deterministic. Decreasing top_p (Option D) narrows the cumulative probability threshold for token selection, which also reduces randomness while still allowing some diversity within the narrowed set. Together, these parameters control the trade-off between creativity and determinism in Amazon Bedrock's text generation.

Amazon Rekognition is a pre-trained service that may not perform optimally for specialized or domain-specific use cases. By using Amazon SageMaker to build a custom model, you can train a model on your own labeled dataset, which directly addresses the mislabeling issue by tailoring the model to your specific images and objects.

Bedrock Guardrails' grounding check ensures the model's response is grounded in the retrieved source documents and can flag or block ungrounded content.

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.

SageMaker Experiments is the correct feature because it is specifically designed to track, organize, and compare machine learning training runs (trials) with different hyperparameters and metrics. It allows data scientists to log parameters, metrics, and artifacts for each run, compare results across runs, and retrieve the exact configuration needed to reproduce the best-performing model.

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.