Free AIF-C01 practice test

Audio recordings and free-text emails have no fixed schema — that's unstructured data. A relational table (B) and a typed CSV (C) are structured. A database index (D) is a lookup structure, not a data-type classification.

The ML lifecycle flows from collecting and preparing data, to training a model, evaluating it, deploying it, and then monitoring it in production (often looping back to improve). The other orderings place steps before their prerequisites exist.

The nesting is: AI is the broadest field (making machines act intelligently); ML is a subset that learns from data; deep learning is a further subset of ML using multi-layered neural networks. The other options invert or scramble this containment.

A large gap between strong training performance and weak performance on new data is the signature of overfitting — the model memorized rather than generalized. Underfitting (B) shows poor performance on both training and new data. Option D contradicts the gap described.

Amazon Q Developer is the generative AI assistant for software development tasks — writing, explaining, and debugging code. Macie (B) discovers sensitive data in S3, QuickSight (D) is BI/dashboards, and Config (C) tracks resource configuration — none is a coding assistant.

LLMs break text into tokens (words or sub-word pieces); usage/pricing and context-window limits are counted in tokens. Option B confuses it with a security credential, and options C and D describe unrelated image and database concepts.

Embeddings are dense numeric vectors that represent the meaning of text/images so that semantically similar items sit close together — the basis of vector search and RAG retrieval. They are not backups (B), model output text (C), or IAM policies (D).

Token counts drive cost (billing is usually per input/output token) and constrain how much text fits within a model's context window. Options A, C, and D are all incorrect — tokens are not about Regions, image color, or authentication.

When a model generates plausible-sounding but false or unsupported content, it's called a hallucination. Overfitting (B) is a training-generalization issue, data drift (D) is changing input distributions over time, and quantization (C) is a model-compression technique.

Encouraging the model to reason through intermediate steps before answering is chain-of-thought prompting, which often improves performance on complex reasoning tasks. The other options are unrelated: zero-shot classification (A), model compression (B), and dataset preparation (C).

A low temperature reduces randomness, producing more deterministic, focused, and consistent outputs — desirable for compliance summaries. Higher temperature increases creativity/variety (B). Temperature affects inference, not training (C) or fine-tuning (D).

RAG grounds responses by retrieving relevant internal documents and injecting them into the prompt at inference time — no model retraining needed, and easy to keep current. Continued pre-training (D) is heavyweight and unnecessary. Temperature (C) and GPU count (A) don't give the model access to internal knowledge.

A vector database stores embeddings and enables similarity search, so a RAG system can retrieve the most relevant chunks to ground the model. It doesn't store model weights (B), replace the model (C), or handle network encryption (D).

Amazon Bedrock Knowledge Bases provides managed RAG — ingesting your data, creating embeddings, storing them in a vector store, and retrieving relevant context for the model. EC2 Spot (A), AWS Backup (C), and Route 53 (B) are compute, backup, and DNS services, unrelated to RAG.

Explainability provides insight into why a model reached a decision, supporting trust, debugging, and compliance. It is not about speed (B), cost (D), or encryption (C), which are performance and security concerns.

Human-in-the-loop review of AI outputs is a recommended safeguard for high-stakes decisions. Full autonomy in high-stakes settings (C) is risky, human review doesn't replace evaluation (B), and responsible AI spans fairness, transparency, safety, and more — not accuracy alone (D).

Bedrock Guardrails let you define denied topics and content filters that block unsafe requests and responses. Max tokens (D), Region (C), and S3 Transfer Acceleration (B) are unrelated to content safety enforcement.

Amazon CloudWatch collects operational metrics and logs and supports alarms and dashboards for monitoring workloads. Artifact (A) is compliance docs, Textract (D) extracts text from documents, and Snowball (B) is data transfer hardware — none is the monitoring service.

Customers are responsible for their data, access control (IAM), and secure configuration/use of the service ('security IN the cloud'). AWS handles the infrastructure — physical servers, datacenters, hypervisor ('security OF the cloud'), which options B, C, and D describe.

IAM defines users, roles, and policies that grant least-privilege access to AWS resources, including Bedrock and SageMaker. CloudFront (C) is a CDN, SNS (B) is messaging, and Global Accelerator (D) is network routing — none is the access-control service.