Evaluation
Validation split
What is validation split?
A validation split is a portion of data held out from training and evaluated periodically during a run, used to detect overfitting and to choose between checkpoints. It informs decisions during development, which is exactly why it cannot serve as the final measurement.
Five to ten per cent is a common allocation. Too small and the metric is noisy enough to mislead; too large and you have given up training data for precision you do not need.
Every decision made by looking at validation performance fits the model a little more closely to that split. This is why a separate held-out set exists, and why using validation numbers as your published result overstates them.
Related terms
Start with the free tier
A magic link creates your account, your tenant and your first API key. No card until you ask for compute.