Training only means something if you can say exactly what it was trained on. A dataset version freezes a set of episodes together with their labels, the train/validation/test split and the preprocessing you chose. It records which recordings the episodes came from and which algorithm versions and settings processed them, so a result can always be traced back and repeated. Versions are numbered per project: v1, v2 and so on.
Create a version
Choose the episodes
Open Versions under your project and choose Create version at the top right. By default the version takes All episodes in this project. To use only some of them, select or filter episodes on Episodes first and choose Create version there. The name fills in with the next number; change it to something your team will recognise.
Under Episodes, two options stay off by default: Include episodes that failed quality checks and Include episodes whose quality checks have not finished. Episodes with warnings are included, and a line tells you how many. If episodes are held back, for example because they are waiting on anonymization or were excluded by hand, the dialog lists the reasons and what to do about each.
Set the split
Under Split, choose a Strategy and the Train share, Validation share and Test share, such as 0.8, 0.1 and 0.1. By recording keeps each recording in one split, so test data never leaks into training. By episode is the better choice when you only have a few recordings, and it is the default for robot projects. You can also split by collector, by task or by manual tags. A Random seed (optional) makes the split repeatable; left empty it uses the version number.
Choose the labels
Under Labels, Latest labels uses the newest labels even if nobody approved them yet. AI-proposed episodes that nobody reviewed are included, and the dialog tells you how many. Approved labels only uses only labels a reviewer approved and holds back episodes nobody reviewed.
Set the preprocessing
Under Preprocessing you can resample to one frame rate, keep only some cameras, cut idle time at the start and end, and cap how long an episode can be. Leave a field empty to keep the data as it is. Then choose Create version. The button stays off until at least one episode is eligible.
The version list
Versions lists every version with its episode count, duration and split counts. Use the ⋯ menu on a row to Set alias, such as prod, or to Move to trash. A trashed version can be restored for 30 days.
The dataset card
Open a version to see its dataset card. It starts with a reminder that the version is frozen, then the episode count, duration, data size and split bar, the verdict summary, the cameras, the tasks and the features with their shapes. Below that is a README you can write with Edit README. On the right are the identifiers you need to cite or load the version (ID, Ref, Digest), the collection source, the privacy state, and who created it and when.
Further down, Use this version gives you copyable Python and CLI snippets and a citation. Export and Train model sit at the top right of the page.
The episodes in a version
The Episodes tab lists what the version holds, each with its split and quality verdict. Use Split to show only train, validation or test.
Versions never change
Once created, a version stays exactly as it is. Later edits to episodes, labels or quality do not reach it, and the dataset card says so. To use different episodes or settings, create a new version. Older versions and their exports stay available, so a model trained last month can still be traced to the exact data it saw.