Domain 3: Applications of Foundation Models

Question 1

Which of the following best describes a key advantage of fine-tuning a foundation model compared to using prompt engineering alone?

The model creates a separate retrieval index from the provided training data The model stores prompt examples in memory for future reference during inference The model gains the ability to access external knowledge bases at runtime The model learns domain-specific patterns embedded in its updated weights

Correct Answer: The model learns domain-specific patterns embedded in its updated weights

Fine-tuning adjusts the model's internal weights using domain-specific training data, allowing it to internalize new patterns and knowledge directly into its parameters, which is fundamentally different from prompt engineering that only guides existing capabilities.

Storing prompt examples in memory describes in-context learning or caching, not fine-tuning. Accessing external knowledge bases at runtime describes Retrieval-Augmented Generation (RAG), not fine-tuning. Creating a separate retrieval index also describes RAG architecture, where documents are indexed for retrieval during inference.

The key concept tested is understanding how fine-tuning permanently modifies model weights to encode domain-specific knowledge, distinguishing it from prompt engineering and RAG approaches.

Question 2

Which classification applies to foundation model evaluation metrics that require a predefined, human-authored baseline to measure the quality of generated text?

Reference-based evaluation metrics Unsupervised alignment assessments Reference-free heuristic validations Autonomous baseline scoring models

Correct Answer: Reference-based evaluation metrics

Reference-based evaluation metrics are correct because they explicitly compare generated text against a predefined, human-authored ground truth or baseline to measure accuracy and quality.

Reference-free heuristic validations are incorrect because they evaluate text outputs using intrinsic properties and linguistic rules rather than comparing them against a human baseline. Unsupervised alignment assessments are incorrect as they evaluate models without relying on labeled datasets or predefined human references. Autonomous baseline scoring models are incorrect because they describe automated evaluation techniques that use another artificial intelligence model as a judge instead of using human-authored text references.

The key concept tested is foundation model evaluation methodologies.

Question 3

What type of model is produced after a fine-tuning job completes on Amazon Bedrock?

A prompt template that adjusts how the original foundation model interprets inputs A retrieval-augmented model that combines the base model with an external knowledge base A custom model that is a specialized version of the base foundation model A completely new foundation model trained from scratch on the provided data

Correct Answer: A custom model that is a specialized version of the base foundation model

When you complete a fine-tuning job on Amazon Bedrock, the result is a custom model that adapts the weights of the original base foundation model using your provided training data, making it a specialized version tailored to your specific use case.

A fine-tuned model does not train a new foundation model from scratch, as that would require massive compute resources and far more data than fine-tuning requires. A prompt template is a separate concept that modifies input formatting without changing model weights. A retrieval-augmented model refers to RAG architecture, which is an entirely different approach that supplements generation with external knowledge retrieval.

Key concept tested: understanding Amazon Bedrock model customization through fine-tuning.

Domain 3: Applications of Foundation Models

AWS AIF-C01 - Domain 3: Applications of Foundation Models Example Questions

Question 1

Question 2

Question 3

Validate Your AI Knowledge on AWS