3.0 Applications of Foundation Models AIF-C01 Practice Quiz

• A. Foundation model selection chooses an FM based on cost, modality, latency, language support, model size, complexity, customization needs, and input/output length.
• B. Model modality describes the type of data a foundation model can process or generate, such as text, image, audio, video, or multiple modalities.
• C. Model latency is the time it takes for a model to return an output after receiving input.
• D. Model size and complexity affect capability, cost, latency, infrastructure needs, and ease of deployment or customization.

Foundation model selection chooses an FM based on cost, modality, latency, language support, model size, complexity, customization needs, and input/output length. This is the correct answer.

• A. Amazon Bedrock Knowledge Bases helps connect foundation models to enterprise data sources for Retrieval Augmented Generation workflows.
• B. Model modality describes the type of data a foundation model can process or generate, such as text, image, audio, video, or multiple modalities.
• C. A vector database stores embeddings so applications can search for semantically similar content.
• D. Amazon OpenSearch Service can store and search vector embeddings for semantic search and RAG use cases.

Model modality describes the type of data a foundation model can process or generate, such as text, image, audio, video, or multiple modalities. This is the correct answer.

• A. Fine-tuning customization tradeoff is the best match when a foundation model application requires this meaning: Fine-tuning updates a model with task-specific examples and can improve behavior for a domain, but requires training data and additional cost.
• B. Foundation model selection is the best match when a foundation model application requires this meaning: Foundation model selection chooses an FM based on cost, modality, latency, language support, model size, complexity, customization needs, and input/output length.
• C. Pre-training customization tradeoff is the best match when a foundation model application requires this meaning: Pre-training trains a model on broad data from scratch or near-scratch and is usually the most expensive customization approach.
• D. In-context learning is the best match when a foundation model application requires this meaning: In-context learning guides a model by including examples or instructions in the prompt instead of changing model weights.

Foundation model selection chooses an FM based on cost, modality, latency, language support, model size, complexity, customization needs, and input/output length. This matches the foundation-model application scenario without confusing model selection, prompting, RAG, customization, training, or evaluation. This is the correct answer.

• A. Negative prompts is the best match when a foundation model application requires this meaning: Negative prompts specify what the model should avoid producing or doing.
• B. Zero-shot prompting is the best match when a foundation model application requires this meaning: Zero-shot prompting asks a model to perform a task without providing examples in the prompt.
• C. Model modality is the best match when a foundation model application requires this meaning: Model modality describes the type of data a foundation model can process or generate, such as text, image, audio, video, or multiple modalities.
• D. Single-shot prompting is the best match when a foundation model application requires this meaning: Single-shot prompting provides one example to guide the model response.

Model modality describes the type of data a foundation model can process or generate, such as text, image, audio, video, or multiple modalities. This matches the foundation-model application scenario without confusing model selection, prompting, RAG, customization, training, or evaluation. This is the correct answer.

• A. Prompt injection is the best match when a foundation model application requires this meaning: Prompt injection attempts to manipulate a model into ignoring instructions, exposing data, or performing unintended actions.
• B. Prompt versioning is the best match when a foundation model application requires this meaning: Prompt versioning tracks changes to prompts so teams can compare, manage, and roll back prompt behavior.
• C. Amazon Bedrock Prompt Management is the best match when a foundation model application requires this meaning: Amazon Bedrock Prompt Management helps create, store, version, and manage prompts for foundation model applications.
• D. Model latency is the best match when a foundation model application requires this meaning: Model latency is the time it takes for a model to return an output after receiving input.

Model latency is the time it takes for a model to return an output after receiving input. This matches the foundation-model application scenario without confusing model selection, prompting, RAG, customization, training, or evaluation. This is the correct answer.

• A. A scenario points to Chain-of-thought prompting when FM selection, prompting, customization, RAG, or evaluation depends on this distinction: Chain-of-thought prompting encourages a model to reason through steps before producing an answer when appropriate.
• B. A scenario points to Prompt templates when FM selection, prompting, customization, RAG, or evaluation depends on this distinction: Prompt templates provide reusable prompt structures that standardize instructions, variables, and context across requests.
• C. A scenario points to Foundation model selection when FM selection, prompting, customization, RAG, or evaluation depends on this distinction: Foundation model selection chooses an FM based on cost, modality, latency, language support, model size, complexity, customization needs, and input/output length.
• D. A scenario points to Prompt guardrails when FM selection, prompting, customization, RAG, or evaluation depends on this distinction: Prompt guardrails define boundaries that help reduce unsafe, unwanted, or off-policy model outputs.

Foundation model selection chooses an FM based on cost, modality, latency, language support, model size, complexity, customization needs, and input/output length. This distinction matters because choosing a nearby concept would lead to the wrong FM design choice, prompt technique, customization method, grounding approach, or evaluation strategy. This is the correct answer.

• A. A scenario points to Domain adaptation when FM selection, prompting, customization, RAG, or evaluation depends on this distinction: Domain adaptation tunes or adapts a model so it performs better for a specific industry, topic, vocabulary, or business context.
• B. A scenario points to RLHF when FM selection, prompting, customization, RAG, or evaluation depends on this distinction: Reinforcement learning from human feedback uses human preferences or ratings to improve model behavior.
• C. A scenario points to Human-in-the-loop evaluation when FM selection, prompting, customization, RAG, or evaluation depends on this distinction: Human-in-the-loop evaluation uses human reviewers to assess model outputs for quality, relevance, safety, or business fit.
• D. A scenario points to Model modality when FM selection, prompting, customization, RAG, or evaluation depends on this distinction: Model modality describes the type of data a foundation model can process or generate, such as text, image, audio, video, or multiple modalities.

Model modality describes the type of data a foundation model can process or generate, such as text, image, audio, video, or multiple modalities. This distinction matters because choosing a nearby concept would lead to the wrong FM design choice, prompt technique, customization method, grounding approach, or evaluation strategy. This is the correct answer.

• A. LLM-as-a-judge is the correct foundation-model application concept when the design, prompt, training, or evaluation decision depends on this exact meaning: LLM-as-a-judge uses a language model to evaluate outputs from another model or application based on defined criteria.
• B. Model modality is the correct foundation-model application concept when the design, prompt, training, or evaluation decision depends on this exact meaning: Model modality describes the type of data a foundation model can process or generate, such as text, image, audio, video, or multiple modalities.
• C. Model latency is the correct foundation-model application concept when the design, prompt, training, or evaluation decision depends on this exact meaning: Model latency is the time it takes for a model to return an output after receiving input.
• D. Foundation model selection is the correct foundation-model application concept when the design, prompt, training, or evaluation decision depends on this exact meaning: Foundation model selection chooses an FM based on cost, modality, latency, language support, model size, complexity, customization needs, and input/output length.

Foundation model selection chooses an FM based on cost, modality, latency, language support, model size, complexity, customization needs, and input/output length. This applies the concept at the point where an AI Practitioner must choose the best foundation-model application decision for the business scenario. This is the correct answer.