AIF-C01 Applications of Foundation Models Applications of Foundation Models Practice Questions

Modality describes the input and output types an FM supports, such as text or images, so evaluating whether a model can output both images and text descriptions is an assessment of modality.

Latency is response speed, context window length is how much input and output fit, and multilingual support concerns languages; none describes the output types a model can generate.

Amazon OpenSearch Service provides a vector database that stores and searches embeddings, which is what powers RAG workflows and integrations with foundation models.

Amazon S3 is object storage, Amazon DynamoDB is a key-value database, and Amazon Redshift is a data warehouse; none offers native vector search over embeddings.

Human-in-the-loop validation with Amazon SageMaker Ground Truth Plus routes labels through expert reviewers who verify and correct them, producing the highest-accuracy annotations with the fewest mistakes.

Amazon Rekognition Custom Labels automates labeling but, without human review, can propagate errors; generating synthetic images with an Amazon Bedrock model creates data yet does nothing to validate the labels; SageMaker Ground Truth's automated active-learning labeling cuts effort but raises the chance of incorrect annotations, the opposite of the goal.

Fine-tuning a model on Amazon Bedrock requires a labeled dataset in which each record pairs an input prompt with the expected completion, so the model learns the desired input-to-output behavior from the company data.

Provisioned Throughput is a capacity and pricing option for inference, not a fine-tuning method. A .txt file containing .csv lines is not the required fine-tuning format. Training on journals and textbooks is unstructured content that lacks the prompt-and-completion pairs that fine-tuning needs.

Iteratively refining the prompt steers the foundation model toward responses that match the company brand tone, because the prompt instructions and examples shape the style of the output.

Limiting output tokens only shortens responses. Batch inference changes how requests are processed, not their tone. Raising the temperature increases randomness, which makes tone less consistent rather than more aligned.