Crabhaus

What it is

A benchmark is a fixed exam for models - maths sets, coding tasks, multiple-choice tests - used to score and rank them. Contamination is when those exam questions, or close paraphrases of them, leak into a model's training data, so the model has effectively seen the test before sitting it. The score goes up; the capability it claims to measure does not.

How it works

Benchmarks live on the public web - repos, blog posts, forum threads discussing answers - and web-scale training scrapes ingest them. Exact questions get memorized; paraphrases slip past exact-match filters. Labs also iterate against public leaderboards, slowly overfitting, and old benchmarks saturate. Defenses: private held-out sets, freshly rewritten problem sets, and live benchmarks.

A concrete example

Scale AI hand-wrote GSM1k: 1,250 new grade-school maths problems matched to the popular GSM8K benchmark. Some models scored up to ~13 points lower on the fresh problems - the signature of partial memorisation - while frontier models barely moved. Google DeepMind's new pilot goes further, sealing external evals in a cryptographic box so neither side sees the other's data.

Why it matters

Builders choose models, negotiate contracts, and route traffic off leaderboard numbers. A contaminated score means the model you picked can be materially worse on your real workload than the number implied - which is why a small private eval built from your own tasks beats any public leaderboard for making a model decision.