Evaluation

LLM evaluation encompasses the methods and metrics used to measure model quality across dimensions including accuracy, safety, instruction following, and reasoning. Evaluation combines automated benchmarks (MMLU, HumanEval), reference-based metrics (BLEU, ROUGE), model-based judging (LLM-as-judge), and human preference studies.

Robust evaluation is essential for guiding training decisions and detecting capability regressions.