Skip to contentTalk to us
العربيةLog inStart free

Evals

In Arabic: التقييمات

AI & language models

Structured test sets that measure the quality of an AI system's outputs in a repeatable, comparable way.

Evals are automated tests for the AI world. You prepare a list of real questions or tasks with expected answers or grading criteria, run the system on them and measure the results. Grading can be exact matching, rules, human review or another model judging the response.

Without evals it is hard to know whether a prompt change or a model switch improved things or made them worse. As models change and get updated, evals help you catch regressions before your customers notice them. Practical tip: start small: 20 to 30 real questions from your customers, including hard cases, different dialects and manipulation attempts. Run them after every significant change and keep the results so you can compare versions.

ExampleBefore switching the clinic assistant to a different model, the team runs it on 30 saved questions about appointments and insurance to confirm accuracy did not drop.
All terms

Related terms

هندسة الأوامرPrompt engineering

The skill of writing, testing and systematically refining instructions for language models so they give accurate, consistent results.

الهلوسةHallucination

When a language model states false or made-up information with full confidence, as if it were fact.

الضوابط الوقائيةGuardrails

Rules and checks placed around an AI model to stop unwanted or risky responses and actions.

اختبار المقارنة أ/بA/B testing

Showing two versions of a page or feature to two groups of users at the same time to see which performs better.

Your idea deserves to see the light.

Describe it in Arabic or English and Fanann builds, tests and ships it — no coding knowledge needed. Start free, no credit card.