Model Evaluation
Model evaluation is the systematic measurement of an AI model's performance, safety, and behavior across representative tasks — using fixed benchmarks, golden datasets, human ratings, and LLM-as-judge methods — both before and after deployment.
Full definition
Classical ML uses metrics like AUC, F1, RMSE on held-out test sets. LLMs require richer evaluation: capability benchmarks (MMLU, GSM8K, HumanEval), safety benchmarks (HarmBench, ToxicChat), domain golden datasets, pairwise preference voting, and LLM-as-judge (Zheng et al., 2023) where a stronger model rates outputs. Evaluation rigor is now a procurement gate: serious enterprise buyers demand task-specific eval suites before contracting.
Why it matters
Without proper evaluation, model selection collapses to vibes. Teams pick the model that "feels best" on five hand-picked prompts and miss systematic regressions. A 200-prompt golden set with quantitative scoring is the minimum viable bar for production AI.
Example
A legal-tech vendor builds a 1,200-question golden set across 12 jurisdictions, scores three frontier LLMs, and discovers the cheapest one wins on jurisdiction-specific accuracy by 7 points.
Related terms
- Eval HarnessAn eval harness is the software framework that runs evaluation tasks against AI models, collects model outputs, applies scoring rubrics, and produces comparable, reproducible metrics across models, prompts, and versions.
- Golden DatasetA golden dataset is a curated, expert-validated set of inputs and reference outputs that serves as the canonical benchmark for evaluating an AI model on a specific task — the team's "ground truth" for is-this-shippable decisions.
- FactualityFactuality is the property of an AI system's outputs being verifiably true with respect to a trusted reference corpus or world knowledge — a distinct dimension from fluency, helpfulness, or generic accuracy.
- AI Red-TeamingAI red-teaming is the practice of systematically probing an AI system — using adversarial prompts, jailbreak techniques, and edge-case inputs — to surface harmful, unsafe, biased, or policy-violating behaviors before and after deployment.
Source & further reading
Primary source: Zheng et al. — "Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena" (NeurIPS) (2023).
Citation policy: this entry is part of the AIDOLS AI Implementation Glossary and may be quoted for research, journalism, and education with attribution to aidolsgroup.com/ko/glossary/model-evaluation/.