AWS Certified AI Practitioner AIF-C01Demo

Amazon SageMaker Model Monitor continuously monitors ML models in production, including model-quality metrics such as accuracy. It detects deviations from configured baselines or thresholds and supports alerts when model performance drifts.

Amazon SageMaker Clarify provides model explainability through feature attribution, showing how individual input features contribute to model predictions. It also supports NLP explainability for text features.

Transformer-based language models use self-attention to model contextual relationships among tokens in an input sequence. This enables each token representation to incorporate relevant information from other tokens.

Once a model is trained, confidential data becomes embedded in the model's parameters. The only way to ensure the model does not generate responses based on confidential information is to delete the model, remove sensitive data from the training dataset, and retrain from scratch.

Content embeddings and the search index are produced during the data-ingestion phase, before user requests arrive. Batch processing these steps keeps newly published content available for subsequent low-latency query embedding, retrieval, and response generation. AWS guidance describes document embedding and vector-database ingestion as the upfront RAG step, followed by query submission, retrieval, and LLM response generation.

BERTScore measures similarity between generated and reference text using contextual BERT embeddings, enabling semantic comparisons. ROUGE measures overlap with the reference and is recall-oriented, making it suitable for assessing whether a summary covers key reference information.

Transfer learning adapts a pretrained model to a new, related task by continuing training with task-specific data, reusing the model’s previously learned representations. AWS documentation identifies fine-tuning of a pretrained model on a new dataset as transfer learning.

Moderation APIs are specifically designed to scan generated content before delivery and filter or block inappropriate material such as explicit or harmful images. They integrate directly into chatbot workflows to prevent harmful outputs. Retraining with public data, model validation alone, and user feedback integration do not provide real-time content filtering capabilities.

Creating a prompt template that instructs the LLM to recognize and resist common attack patterns (prompt injection, malicious requests) strengthens the model's ability to defend itself against manipulation. By explicitly teaching the model its boundaries and expected behaviors, you reduce the risk of undesirable outputs and data exposure. Increasing temperature amplifies randomness (counterproductive), restricting to specific platforms lacks security benefit, and limiting input tokens risks losing critical context.

Sentiment analysis extracts or classifies opinions, attitudes, and emotional tone from human language, making it a natural language processing (NLP) application.

Embeddings encode items as numerical vectors in a reduced-dimensional space, allowing models to represent and compare semantic or structural relationships mathematically.

The prompt attacks policy detects and filters jailbreaks and prompt-injection attempts intended to override developer instructions or bypass a foundation model’s built-in safety and moderation capabilities.

Amazon SageMaker Clarify's primary functionality is identifying potential bias in datasets during data preparation, helping practitioners detect imbalances and fairness issues before model training. This early detection is crucial for building responsible and fair ML systems. RAG workflows are for retrieval-augmented generation (a different task), Model Monitor (not Clarify) monitors production performance, and model documentation is a different function.

Amazon Personalize generates personalized recommendations from users’ item-interaction history. Its User-Personalization-v2 recipe is specifically suited to personalized movie recommendations for a streaming application.

Temperature is the inference parameter that controls response randomness. A temperature of 0 minimizes randomness and therefore yields the most deterministic and stable output for the same prompt; instructing the model in prompt text does not provide the same sampling control, while a temperature of 1 increases variation.

AWS PrivateLink enables private, secure connectivity between VPCs and AWS services without exposing traffic to the public internet. This satisfies regulatory compliance requirements that prohibit internet access while allowing the VPC to access Bedrock and other AWS services over AWS's internal network. Macie provides data security, CloudFront distributes content, and an Internet Gateway provides public internet access—none solve the compliance constraint.

ROUGE (Recall-Oriented Understudy for Gisting Evaluation) is specifically designed to measure how well generated text matches reference examples by calculating n-gram overlap, precision, recall, and F1 scores. Since the company has reference examples of readable text and wants the model's output to resemble them, ROUGE directly evaluates this alignment. The loss function measures training performance, semantic robustness measures resilience, and latency measures speed—none assess output similarity to examples.

Amazon Bedrock models can have different maximum generated-response token limits. A prompt and its requested output length must therefore remain within the maximum token count supported by every target model. AWS notes that inference-parameter defaults and valid ranges depend on the model, and maxTokens defaults to the maximum permitted by the selected model.

A text embeddings model converts text into numerical vector representations that capture semantic meaning. Those embeddings can be stored in Amazon OpenSearch Service for vector similarity search and retrieval of relevant private data. Amazon Titan Text Embeddings models

The context window defines the maximum number of tokens (input and output combined) a foundation model can process in a single prompt. This directly determines how much information can fit into one prompt. Temperature controls output randomness, batch size relates to throughput efficiency, and model size refers to parameter count—none of these affect individual prompt length limits.

Adjusting the prompt allows direct control over LLM output characteristics including length and language. By crafting instructions like 'Respond in [language] with 2-3 sentences', the model aligns with requirements. Model size, temperature, and top-K control different aspects of generation (creativity, diversity) but do not directly constrain output length or language.