Decoder. plain-English AI glossary

Contamination check

● Core

Scanning your training data for examples that appear in public benchmarks, which would invalidate your results.

Think of it like

Checking whether your school essay was plagiarized from Wikipedia.

Example

You train a model on Common Crawl. Later you discover your training data includes 1% of the MMLU benchmark. Your MMLU score is no longer credible.

How it actually works

Large-scale training makes contamination easy — you're training on billions of tokens and those tokens came from the internet, which includes benchmarks. Major labs now report contamination percentages. Fixing it requires deduplicating against benchmarks (expensive) or using evaluation code in training data and being honest about it (rare).

For product teams

Benchmark scores without contamination analysis are meaningless. Demand evidence of decontamination.

For engineers

Download public benchmarks. Deduplicate training data against them using hashing and embedding similarity.

Integrity check.

Read anything AI without the jargon

Look up any term in plain English, or save terms as you read with the free Chrome extension.

Open Decoder Add to Chrome