LLM evaluation

LLM evaluation is the set of methods used to measure how well a language model performs on the tasks it will actually be asked to do, whether it is being chosen off the shelf, compared after fine-tuning, or checked before a release. It sits alongside, but is distinct from, LLM observability, which watches a model's behavior after it is already in production.

Evaluation combines several approaches: automated benchmarks with a known right answer, human review against a rubric, and increasingly LLM-as-a-judge, where a separate model scores outputs for qualities like correctness, tone, or safety. Good evaluation sets are built from realistic examples, including edge cases and known failure modes, rather than only easy questions the model is likely to get right.

Evaluation matters because a model's fluency is not evidence of accuracy; without it, teams cannot tell whether a change made a system better or introduced regressions such as new hallucination patterns or unsafe outputs. Common pitfalls include evaluating on the same data used for tuning, which inflates scores, relying on a single aggregate metric that hides failures on specific inputs, and skipping re-evaluation after a model or prompt update.