Back to glossary

Term

Evals

Evals are systematic tests for LLM applications — fixed test cases, automatic scoring of the responses, result as a score. The foundation for objectively comparing prompt or model changes.

Evals — explained in more detail

Anyone building an LLM application knows the problem: a tiny prompt tweak improves one example and breaks three others — without test coverage, that stays invisible. Evals solve this by running a curated set of input/expectation pairs against the system and scoring each response. Scoring can be deterministic (exact match, regex), structural (valid JSON schema?), via LLM-as-a-judge (another model assesses) or hybrid. Individual scores aggregate into a single number that is compared across builds.

Example / Practical context

OpenAI Evals, Promptfoo, Langfuse, Braintrust and LangSmith offer frameworks for this; Anthropic provides its own tooling with anthropic-evals and the Workbench. A typical dataset for a support bot: 50–500 sample tickets with expected tags, tonality criteria and escalation flags. CI pipelines can run eval suites on every pull request and block a merge on score regression.

Public benchmarks like MMLU, HumanEval or Terminal-Bench measure general model capabilities — evals, by contrast, test the concrete application with its own prompt stack. Unit tests deterministically verify a function; evals tolerate variance in the response and measure statistically across many examples. Observability tools collect live data that often feeds back into eval datasets.

Entdecke mehr

Themenuebersicht