A dataset version is a frozen set of episodes, with a train, validation and test split and the preprocessing you chose. Each project numbers its versions v1, v2, and so on. Export a version to download it, or train a model on it.
Who can do this: Owner, Editor (project roles).
Before you start
- Episodes exist in the project. Recordings must be processed and split into episodes first.
Open Versions
In the sidebar, open your project and choose Versions. The list shows every version with its episode count, duration and split counts.
Open Create version
Click Create version at the top right. The name fills in with the next free number, such as v2. Change it if you want.
Choose the episodes
Under Episodes, leave Include episodes that failed quality checks and Include episodes whose quality checks have not finished off unless you need them. Both are off by default. Episodes that carry warnings are included. The line under the options shows how many episodes and how much time the version will hold.
If you started from Episodes with some episodes selected, the version uses only those episodes.
Choose the split
Under Split, choose a Strategy and set the Train share, Validation share and Test share. Enter a Random seed (optional) to fix the split. The seed defaults to the version number. The strategy decides how episodes are assigned: by episode, by recording, by collector, by task, or by manual tags. Robot projects start at By episode, human projects at By recording.
| Field | What to enter | Example |
|---|---|---|
| Strategy | How episodes are assigned to the splits. | By episode |
| Train share, Validation share, Test share | The fraction of episodes for each split. Enter zero or more. The defaults are 0.8, 0.1 and 0.1. | 0.8 |
| Random seed (optional) | The same number always gives the same split. Leave it empty to use the version number. | 42 |
Choose the labels
Under Labels, choose Latest labels to use the newest labels, even if nobody has approved them. Choose Approved labels only to use only the labels that a reviewer approved.
Set the preprocessing
Under Preprocessing, set the frame rate, the cameras to include, and whether to cut idle time. Leave a field empty to skip it.
| Field | What to enter | Example |
|---|---|---|
| Frame rate (fps) | Resamples every recording to this rate. Leave it empty to keep each recording's own rate. | 30 |
| Cameras | Camera names separated by commas. Leave it empty to include all cameras. | head, wrist |
| Cut idle time at the start and end | Removes stretches where the robot or person is not moving. | — |
| Longest episode (seconds) | Episodes longer than this are shortened to it. Leave it empty for no limit. | 60 |
| Margin around the cut (milliseconds) | Extra time kept on each side after the idle time is cut. Only used when idle time is cut. | 0 |
Choose Create version
Check the summary and the warnings. The line above Split shows how many episodes the version will hold, and it updates as you change the options. Then click Create version at the bottom of the dialog.
Read the version page
Each version has its own page with the title, the episode count, the duration and the creation time. Use the buttons at the top to export or train.
| Area | What it does |
|---|---|
| ① Export | Opens the export dialog. See Export and download. |
| ② Train model | Starts training on this version. See Training. |
| ③ More actions | Opens Set alias. An alias is a name you can use for the version. The names latest, sha256, head, main and new are reserved. |
| ④ Card, Episodes, Files | Card: the summary, README, how to use the version, and its exports. Episodes: the episodes in the version, filtered by split. Files: the files of the version, with size and SHA-256. |
A version does not change
The episodes, labels, split and preprocessing of a version are fixed when you create it. The page says so: "This version is frozen. Later episode edits do not change it." To use different episodes or settings, create a new version. Older versions and their exports stay as they were.
Troubleshooting
Create version is disabled
You need permission to edit data in this project, which only the project Owner or Editor roles have. Ask one of them. The button also stays disabled until at least one episode is eligible.
The dialog says no episodes can go into the version
The dialog lists why episodes were left out, for example failed checks, unfinished checks, missing consent or anonymization, label revisions that nobody reviewed, or episodes excluded by hand. Fix those episodes first. If some of them are acceptable for this version, turn on the matching option.
The list is empty
Episodes appear after a recording has been processed and split. Until then there is nothing to put in a version.
Can I delete a version?
Choose Move to trash from the row menu on the list. You can restore the version for 30 days. Storage is still billed until the version is permanently deleted.