Benchmark
A shared public test used to compare models.
A standardised task set with published scores — MMLU, HumanEval, SWE-bench, GPQA — used to rank models. Useful for coarse comparison, poor at predicting performance on your specific problem.
Benchmarks suffer from contamination: popular test sets can end up in training data. Treat published scores as a first filter, never as a decision.
Use benchmarks for coarse triage — ruling a model in or out — and never as the decision. The gap that matters is between a shared score and your task distribution, and that gap is usually wide, because your task is specific, messy and unlike the benchmark’s construction.
The leaderboard leader that lost
The top-scoring model is selected and underperforms the previous one in production.
Its edge was on reasoning benchmarks; the actual workload is structured extraction from messy documents, where a cheaper model with schema-constrained decoding did better. The eval set drawn from real inputs would have shown that in an hour.
Scores that may be partly memory
A model jumps several points on a well-known public benchmark.
Public test sets leak into training data. Treat a suspicious jump as a prompt to check for contamination before citing it, and prefer private or freshly gathered sets for decisions that matter.
- paperMeasuring massive multitask language understanding ↗
MMLU across 57 subjects. Read it to know what the benchmark tests — not as evidence that one number summarises capability.
- paperInvestigating data contamination in modern benchmarks for large language models ↗
Detecting test-set overlap between training data and benchmarks, and why a high score on a public benchmark may be partly recall of the answers.
- blogA statistical approach to model evaluations ↗
Why a benchmark-score difference may be noise: report uncertainty, account for clustered questions, use paired comparisons.