What an Eval Actually Is

Evals

Part of: Fine-Tuning & Evals

You tweaked the prompt. Output feels better. Does it? You have no idea, and "feels better" is exactly how teams ship regressions with a straight face. One engineer "improves" the system prompt, three customer-facing answers quietly get worse, and nobody notices for a week. An eval is the fix: a repeatable test that turns a vague hunch into a number. What it is An eval is a unit test for an AI system. It has three parts: 1. A dataset : inputs paired with what good looks like (the reference, or ground truth ). 2. A task : run each input through your model + prompt. 3. A scorer : compare the model's output to the reference and produce a score. Run it, get a number like "87% correct." Change the prompt, run it again, see if the number went up or down. That's the whole game. Without it you're flying blind, trusting your memory of a handful of outputs to represent the whole distribution. How it works A normal unit test checks add(2, 2) == 4. Exact match, one right answer. AI outputs are fuzzy, "The capital is Paris." and "Paris" are both correct, so your scorer has to be smarter than ==. You pick a scoring method that fits the task: - Exact match : classification or structured fields whe

Challenge: The Regression Gate