Skip to content
calwm

Applications of foundation models cheat sheet

All 10 AIF-C01 flashcards for Applications of foundation models on one page, grouped by exam objective. Print it or save it as a PDF for a last-minute read.

Drill these as flashcardsRead the Applications of foundation models guide

Describe design considerations for applications that use foundation models

Customising a model: cheapest to most expensive?
Prompt engineering, then RAG, then fine-tuning or continued pre-training. Building your own foundation model costs the most.
Where can Amazon Bedrock Knowledge Bases store vectors?
In a vector store such as Amazon OpenSearch, Amazon Aurora PostgreSQL, Amazon Neptune Analytics or Amazon S3 Vectors. It holds the embeddings so matching chunks are found fast.
What does prompt caching save?
Time and input-token cost. When many requests reuse the same long context, such as a document, the model doesn't reprocess it every time.

Choose effective prompt engineering techniques

What is chain-of-thought prompting?
Asking the model to break a complex question into smaller, logical steps. It helps with problems that need reasoning, not just recall.
What is a negative prompt?
A list of what you don't want in the output, kept separate from the main prompt. For example, "waves, clouds" to keep them out of a generated image.
Prompt injection vs jailbreak?
Prompt injection tries to override the developer's instructions, as in "ignore everything earlier". A jailbreak tries to get past the model's own safety rules to produce harmful content.

Describe the training and fine-tuning process for foundation models

Fine-tuning vs continued pre-training?
Fine-tuning uses labelled prompt-and-answer examples to teach a task. Continued pre-training uses large amounts of unlabelled text to teach a domain's language.
What is model distillation?
Transferring knowledge from a large "teacher" model to a smaller "student" model. You get a faster, cheaper model for your use case.
What is RLHF?
Reinforcement learning from human feedback: people rate the model's responses, and those ratings guide its training. It aligns outputs with what people actually want.

Describe methods to evaluate foundation model performance

How do you show an FM app is worth the money?
Track business KPIs, not just model scores: cost per interaction, hours saved, revenue lift and customer satisfaction. Together they show the return on investment.