Eval (evaluation)
- Checking AI's work
- Standard term
An eval, short for evaluation, is a repeatable test of an AI system. It pairs a saved set of cases with criteria for a good result and a way to score or review each one.
AI answers vary from run to run and shift with each model or prompt change, so a few questions tried by hand say little. An eval holds the cases and scoring fixed, so a score change comes from the system or from chance. Enough cases, sometimes run more than once, keep chance small.
Answers with one right value, like a category or a total, can be scored by code. Open-ended ones, like the tone of a reply, need a written rubric, a person comparing outputs side by side, or another AI acting as the grader.
Where you’ll meet it
Your team uses an AI feature that sorts incoming customer emails into billing, bugs, and everything else. A vendor announces a newer model and says it's more accurate. A coworker wants to switch today, and you remember the current one filing a refund request under everything else last month.
What to do
- Gather a hundred or more past emails, tricky ones and that refund request included, and label each one correctly.
- Ask whoever set up the feature to run both models on that same set.
- Read the emails each model got wrong before deciding, since a gap of a point or two can be chance.
Easy to mix up
A benchmark compares models from many labs on one shared public set of questions. An eval compares choices for your own system, on cases drawn from your work.
Check yourself
Your company's help-desk chatbot scores 95% on an eval built a year ago, the same as it did then. This quarter, customer complaints about its answers have doubled. Which is the more likely explanation?
Go deeperAI graders and leaked test sets can both inflate a score
An AI grader is itself a model, with biases of its own. A 2023 study of models as judges, by Zheng and colleagues, found they tended to favor the answer shown first, longer answers, and answers written by their own model. Teams check a grader by having people score a sample too and measuring how often the two agree.
Public benchmark questions circulate online, and a model trained on text that includes them can score well by recall. This is called “benchmark contamination.” A set that never leaves the team is unlikely to end up in training data, which is one reason teams don't publish the sets they rely on.