Canary
Also called Canary String
A unique marker planted in data so you can later tell if it leaked or got memorized.
Think of it like
A dyed banknote in the till — if it shows up somewhere it shouldn’t, you know exactly where it came from.
Example
A benchmark embeds a rare string in its questions; if a model can recite it, you know the benchmark contaminated its training data.
How it actually works
Canaries are unique, unlikely-to-occur tokens dropped into documents, datasets, or benchmarks. Finding one where it should not be proves leakage — training-data contamination, a data exfiltration path, or unauthorized copying. Named for the canary in a coal mine: a cheap early-warning signal.
For product teams
A near-free tripwire that tells you when your test set has quietly leaked into a model.
For engineers
Embed unique tokens to detect memorization, benchmark contamination, or exfiltration; presence downstream is proof of leak.
Related
- Tripwire — A specific kind of tripwire.
- Evaluation gaming — Helps detect the contamination behind evaluation gaming.
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