Train-Test Leakage
Also called Data Leakage
When test answers sneak into training, so the model looks brilliant on a test it has secretly already seen.
Think of it like
Grading a student on questions they were accidentally handed the night before.
Example
A web-scraped training set turns out to contain the exact benchmark questions, and the reported score is meaningless.
How it actually works
Leakage happens whenever information from the test set — the exact examples, or features derived from them — is present at training time. The model memorizes rather than generalizes, so evaluation numbers are inflated and collapse in the real world. With web-scale training data, accidental contamination of public benchmarks is a constant, hard-to-detect risk.
For product teams
A leaked benchmark tells you nothing; always ask whether the eval data could have been in training.
For engineers
Contamination of the evaluation set into training data, producing optimistic and non-generalizing metrics.
Related
- Overfit to Eval — A cause of overfitting to eval.
- Data contamination — Related to dataset contamination.
- Holdout — Why you hold out clean data.
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 DecoderAdd to Chrome