Think Build Implement Repeat
London, UK +44 7367 067226
WhatsApp FOLLOW f in X
Python & Django

Knowing Whether a Change Made Things Better

Last updated:

You cannot improve what you cannot measure

Prompt engineering without an evaluation set is adjusting text and forming impressions. It feels productive and it produces regressions nobody notices.

The single most valuable artefact in an AI project is a hundred real cases with agreed correct answers. Everything else follows from having it.

Building the set

  1. Take real cases from your actual data or support log
  2. Have someone who knows write the correct answer for each
  3. Include cases where refusal is correct — questions with no good answer
  4. Include ambiguous cases where asking back is correct
  5. Hold back a portion you never tune against

How to score

TaskScoring
ClassificationExact match against the label
ExtractionField by field, exact or normalised
AnsweringModel-assisted grading, with human-checked samples
Refusal casesDid it decline when it should?
RetrievalWas the right passage in the top results?

Extraction and classification score cleanly. Free-text answers need model-assisted grading with periodic human verification of the grader itself.

Run it automatically

  • On every prompt change
  • On every model change
  • On a schedule, to catch provider-side changes
  • With the score recorded over time, so trends are visible
  • Blocking a deployment if it drops beyond normal variation

Refresh it

A set built in January describes January. Add new real cases monthly, especially failures, and retire cases that no longer represent your work.

Without refreshing, you eventually optimise for a test set that no longer resembles reality, which is worse than having none.

Frequently asked questions

How many cases do we need?

A hundred is a workable minimum; two to three hundred gives more stable scores. Representative matters more than large.

Who writes the correct answers?

Someone in the business who knows. This is the part that cannot be outsourced.

Can a model grade itself?

Model-assisted grading works for scale, checked against human-verified samples. Purely self-graded evaluation tells you very little.

How long does building one take?

Two to four days with a domain expert. It is the highest-value time in the whole project.

Keep reading

Adjusting prompts and hoping?

An evaluation set turns that into a measurement. We can build one around a system we did not write.

Book a free 30-minute call Get a project estimate WhatsApp us

Related services

What we build for problems like this one

Custom Software DevelopmentWeb DevelopmentMachine Learning