Decoder. plain-English AI glossary

Holdout

● Core

Also called Holdout Set

Any chunk of data you deliberately set aside and keep the model from seeing, so you can test it fairly.

Think of it like

Hiding a few puzzle pieces so you can check later whether someone actually solved it or just memorized the picture.

Example

Before training, the team carves off 10% of records as a holdout and does not look at it again until the end.

How it actually works

Holdout is the general idea behind validation and test splits: withhold data so evaluation reflects performance on examples the model never learned. Simple holdout uses one fixed slice; when data is scarce, cross-validation rotates the holdout across the whole set to use every example both ways.

For product teams

The basic discipline that keeps a demo honest — no holdout, no trustworthy number.

For engineers

Data withheld from training for unbiased evaluation; the fixed-split alternative to cross-validation.

Related

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