AI evals

Sefism.AI evals

Tauseef Fayyaz

Chapter 01

What evals are

The parts every eval has, and why judging changes by feel stops working after the first week.

Eval

Also called evaluation, LLM eval

What: A repeatable test of an AI feature: a fixed set of inputs, a way to grade each output, and a score you can compare between versions.
Why: Every prompt, model or retrieval change becomes an experiment with a before and after, instead of an opinion.

Dataset, task, grader, score

What: The dataset holds cases; the task is your app or prompt producing an output; the grader decides pass, fail or a score; the score is aggregated across cases.
Why: Most tools use these four words in some form, so once you see them every framework looks familiar.

Python

for case in dataset:
    output = run_app(case["input"])
    results.append(grade(case, output))
score = sum(r.passed for r in results) / len(results)

Evals vs unit tests

What: Unit tests expect one exact result. Evals accept that outputs vary, so they report a pass rate across many cases and often several runs.
Why: Expecting 100% on an LLM eval makes the suite flaky; set a threshold and watch the trend.

Public benchmarks vs your evals

What: Benchmarks score models on shared tasks; your evals score your feature on your users' tasks.
Why: A model that leads a leaderboard can still fail your support questions in Urdu or your JSON format.

Eval-driven development

What: Write a handful of cases before changing the prompt, run them, make the change, run them again.
Why: It is test-driven development adapted for outputs that are not exact.

Go deeper