Skip to main content
A dataset version (or snapshot) is a frozen, point-in-time copy of your labeled data with train/validation/test splits. It’s what you train on — and because it never changes, the exact same run can be reproduced later.

​
Set your splits

In the Dataset view, choose how images are divided into train, validation, and test. A common split is 70 / 20 / 10. Images you’ve assigned manually keep their split; the rest are distributed to hit your target ratios.
Dataset split

Distribute images across train, validation, and test.

​
Create a snapshot

  1. Open Snapshots and select Create Snapshot.
  2. Optionally add preprocessing and augmentation steps — see Augmentation.
  3. Confirm to create the version.
AcuSight freezes the current dataset into a numbered version (v1, v2, …) you can train on.
Dataset versions are immutable — once created, their contents never change. New labels or batches mean a new version. Each version trains one model, so “which data produced this model” is always unambiguous.

​
Next steps

Augment your training data

Add preprocessing and augmentation to get more out of your images.

Model lifecycle

How a dataset version becomes a deployed model.