MEPX
Chapter 6 of 10All chapters

Chapter 6 of 10

Customising a model

Three levels of effort.

The options

Prompting shapes behaviour per request. Retrieval supplies documents at run time. Fine-tuning adjusts weights on your own examples. Cost and effort rise steeply in that order.

  • Most problems that look like they need fine-tuning need better retrieval.
  • Fine-tuning teaches style and format more reliably than it teaches facts.

Evaluating

Write a set of representative inputs with expected qualities before changing anything. Without that, comparisons between prompts or models are impressions rather than evidence.