AI Learn Home
Evaluation

How to evaluate a model for your own task

11 September 2026 · 7 min read

Public leaderboards answer a question you do not have: which model is best on a mixture of tasks chosen by someone else. The question you do have is narrow, and it needs its own data.

Build a golden set first

Fifty examples is enough to start. Not fifty random samples — fifty that represent the distribution you actually see, including the ugly cases: the empty input, the contradictory request, the document in the wrong language.

For each example record the expected output in a form you can check. Exact matches are rare in generation, so most items will need either a rule (does it contain the account number?) or a human label (is this summary faithful?).

Score the components, not only the answer

Production failures usually happen at a step, not at the end. If your system retrieves, extracts and then writes, measure each stage separately. A final-answer score of 78% tells you nothing about which of the three stages to fix.

Pairwise beats absolute scoring

Asking “rate this answer from 1 to 10” produces a distribution that drifts between runs and between annotators. Asking “which of these two answers is better, given the rubric” is far more stable, and it maps directly to the decision you are making: keep the current model or switch.

A practical setup: generate answers from both candidates on the same inputs, randomize the order, have a human (or a judge with a tight rubric) pick a winner, count. Two hundred comparisons give a usable signal; twenty do not.

Using a model as a judge

It works, with caveats that are easy to state and easy to ignore:

Measure what the product cares about

Accuracy is only one axis. Also record:

A model that is 3% more accurate and twice as slow is usually the wrong choice, and no leaderboard will tell you that.

Keep the set alive

Add every production failure to the golden set. Re-run the whole suite on every model change, every prompt change and every retrieval change. The set becomes the institutional memory of what “good” means for your product — and the only defence against a confident regression.