AI glossary · Basic · Data & evaluation
Evals
Evals (evaluations) are test sets, and the process of using them, that repeatedly check whether an AI model or AI feature works as intended, using predefined problems and scoring criteria. Unlike public benchmarks, they are usually built in-house to fit your own service and tasks.
About this term
- Level
- Basic · Technical terms you often meet in AI news
- Category
- Foundations & applications
- Fields
- Data & evaluation
- Added
- 2026-10-08
How it works
First, collect inputs that resemble real usage (questions, documents, task instructions and so on) to build a list of test questions. Each question comes with its correct answer or the criteria for a 'good answer.' Then the model or AI agent works through these questions, and the results are graded. Grading may compare outputs directly with the correct answer or check them automatically with code, and it may also mix in LLM-as-a-Judge, where another LLM does the grading, or review by people. The resulting scores are measured again every time the prompt or model changes, to see whether things got better or worse.
Why it matters
An LLM can give different answers to the same input, and changing a single line of a prompt can cause performance to drop unexpectedly somewhere else. Without Evals, judgments end up resting on a feeling that 'it seemed fine after a few tries.' With Evals in place, decisions such as switching models, editing prompts or changing RAG settings can be weighed with numbers and examples. For this reason, many teams building AI products treat Evals as a basic development tool, much like test code in conventional software.
Things to keep in mind
A high score does not necessarily mean a good experience for real users. If the test questions do not represent real situations well, you end up measuring the wrong thing. And if test questions leak into a model's training data, data contamination can inflate scores beyond the model's actual ability. When an LLM does the grading, the grading model itself can be wrong or biased, so it is safer to have people review a portion directly to confirm the grading criteria are working. As a service changes, its Evals need to be updated along with it.
Example
A company building a customer support chatbot might collect several hundred test questions based on real inquiries, such as 'questions about the refund policy,' 'complaints about delayed delivery' and 'questions it cannot answer.' Each question is given grading criteria such as 'Did it answer based on the policy documents?', 'Did it avoid making up things it does not know (hallucination)?' and 'Is the tone polite?' Every time the team edits the system prompt or tries to switch to a cheaper model, it runs these Evals to compare scores, and if any item drops, it finds the cause before deployment. Companies that build AI models also check performance and safety with their own Evals, alongside public benchmarks, before releasing a new model.