Free AIF-C01 practice test: 30 questions with answers

Predicting a continuous numeric value from labeled historical data is regression. Clustering groups unlabeled data, reinforcement learning learns from rewards, and anomaly detection flags unusual points rather than predicting a value.

Unsupervised clustering groups unlabeled data by similarity, which reveals natural topic groups. Classification needs labels, regression predicts numbers, and reinforcement learning needs an environment with rewards.

Overfitting means the model memorizes training details and generalizes poorly to new data. Equal performance indicates good generalization, underfitting describes overly simple models, and training speed is unrelated.

The test set is held out until the end to give an unbiased estimate of generalization. Training data fits parameters, validation tunes choices, and storage is unrelated to evaluation.

Amazon SageMaker is the managed service for the full machine learning lifecycle, including building, training and deploying models. Rekognition provides prebuilt vision APIs, Polly converts text to speech, and Budgets tracks costs.

Amazon Textract extracts text, tables and forms from documents using managed models. Transcribe converts speech to text, Translate converts language, and Lex builds conversational interfaces.

RAG retrieves current passages from the document store and adds them to the prompt, so answers follow updated documents without retraining. Annual fine-tuning lags behind changes, high temperature increases randomness, and manual prompting does not scale.

Knowledge Bases for Amazon Bedrock manage ingestion, embeddings and retrieval for RAG. Guardrails control content, custom model import brings weights in, and batch inference runs jobs but does not connect data sources.

Fine-tuning on labeled domain examples adapts the model to specialized responses. Using no domain data, renaming the model and removing context do not improve domain behavior.

Instruction tuning trains on instruction and response pairs to improve following tasks. PCA reduces dimensionality, sweeping tunes hyperparameters, and scaling is a preprocessing step.

Evaluating with representative prompts and measuring quality, cost and latency gives evidence for the choice. Size alone does not guarantee quality for a task, and names or single prompts are not evaluations.

ROUGE measures overlap between generated and reference text, which is common for summarization evaluation. Pixel error, frame rate and packet loss are unrelated to text quality.

Foundation models are large, pre-trained models that can be adapted or prompted for many tasks. Single-task models, datasets and rule-based systems do not match this definition.

Tokens are the units models read and generate, and they can be whole words or parts of words. They are not keys, and although token counts affect billing, a token is not a billing unit alone or a layer.

The context window is the amount of input and prior output the model can attend to in one request. Learning rate and epoch count are training settings, and batch normalization is a layer technique.

Temperature scales the probabilities used for sampling, so higher values make less likely tokens more probable and increase variety. It does not change token limits, retrain models or enlarge context windows.

Few-shot prompting includes examples in the prompt to guide the model's behavior without changing its weights. Distillation, gradient clipping and feature hashing are different techniques.

Hallucinations are plausible-sounding outputs that are not grounded in facts or sources. Crashes, temperature settings and refusals are different behaviors.

IAM policies restrict which principals can invoke Bedrock models, which enforces least privilege. Public endpoints, shared keys and disabled audit trails weaken security and governance.

CloudTrail records API calls including model invocations for auditing. Polly, Transcribe and Textract are AI services themselves and do not provide audit trails.

Amazon Macie discovers and classifies sensitive data such as personal information in S3. Lex, Forecast and Snow are not data discovery services for S3.

Encryption at rest and in transit with managed keys and access controls protects sensitive training and inference data. Unencrypted storage, plain HTTP and shared keys in documents expose it.

AWS Config records configurations and evaluates them against rules, providing continuous compliance evidence. Comprehend, Personalize and Snowcone do not evaluate configuration compliance.

AWS Artifact provides compliance reports and agreements that support audit evidence. Kendra is search, Cloud9 is an IDE and Lightsail is simple compute.

Different outcomes for similar profiles across groups indicates bias or unfairness. Drift refers to changing data over time, latency is a performance measure, and residency is a data location requirement.

SageMaker Clarify detects bias in datasets and models and explains predictions. Feature Store manages features, Ground Truth labels data, and Edge Manager deploys models to devices.

Bedrock Guardrails define policies such as denied topics, content filters and sensitive information handling for inputs and outputs. Batch inference runs jobs, Trusted Advisor gives best-practice checks, and Inspector scans for vulnerabilities.

Model cards communicate intended use, data, evaluations and limitations, which supports informed decisions. Private-only documentation, hidden results and removed data information reduce transparency.

Content filters, human review for high-risk outputs and ongoing monitoring form a layered approach to responsible use. Trusting the provider alone, publishing unreviewed and uptime-only monitoring leave gaps.

Explainability means that people can understand the factors behind a model's decisions. Elasticity, replication lag and multi-tenancy are cloud architecture concepts.