A company is comparing methods for using a general-purpose foundation model adapted to its own operations. Distinguishing them from infrastructure optimization, which are representative methods for adapting a pretrained foundation model to the company's specific use case? (Choose TWO.)
A question about choosing TWO methods for adapting a foundation model to a use case.
Prompt engineering, which refines the instructions and examples provided
Correct. Prompt engineering is the most lightweight adaptation method, which improves output by refining the instructions and examples provided without changing the model's weights.
Fine-tuning, which performs additional training on company data
Correct. Fine-tuning is an adaptation method that performs additional training on the company's labeled data to update the model's weights and specialize it for a specific use case.
Scaling out the inference endpoint
Scaling out is an infrastructure response that increases processing capacity.
It increases the volume that can be handled, but it is not a method for adapting the model's responses to the company's use case, so this is incorrect.
Caching and reusing output
Caching is a refinement that speeds up and lowers the cost of responses to the same question.
It is not a method for changing the content of responses to suit the company, so this is incorrect.
Changing to a Region closer to users
Changing the Region is a response to latency or location requirements.
It is not a method for adapting the content of the model's responses to the use case, so this is incorrect.
There are two representative methods for optimizing a foundation model for a company.
- Prompt engineering: without changing the model's weights, refine the instructions and examples (few-shot) provided to steer output toward the company's use case. Its advantage is that it can be tried quickly and cheaply.
- Fine-tuning: perform additional training on the company's labeled data to update the model's weights and specialize it for a specific operation or domain. It takes effort and cost but can raise expertise and accuracy.
First try the lightweight prompt engineering, and if that is not enough, move to fine-tuning (others include RAG and continued pretraining).