Bad data costs you a training run. Every recording and episode is checked automatically as it is processed, and the Quality page puts the results for the whole project in one place, so you can spot problems before you build a dataset version.
What is checked
The checks depend on the kind of data. Robot recordings are checked for things like timestamp and clock problems, gaps between streams, jumps, frozen or saturated actions, and values that are not numbers. Human first-person recordings are checked for blur, underexposure, how often the hands are in view, idle time, and whether any face may have been left unblurred after anonymization. Some checks also look across the whole project, for example duplicate episodes or a split that leaks between train and test.
How to read a result
Each check ends with one of six results. Only Pass means the check passed, and only Pass is shown in green. A warning, a failure, a check that could not finish, a check that has not run yet and a check that does not apply to this data are all something other than a pass, and the page never shows them as passed.
| Result | What it means |
|---|---|
| Pass | The check ran and passed. |
| Warning | The check found something to look at. The data can still be used. |
| Fail | The check failed. |
| Error | The check could not finish. |
| Not checked yet | The check has not run. |
| Does not apply | The check does not fit this data, for example a hand check on a robot recording. |
Each episode also gets an overall verdict: Pass, Warning, Fail, Incomplete (some checks did not run or could not finish) or Not checked. You see it on every episode card and can filter by it with Verdict.
The project quality page
Under your project in the sidebar, choose Quality. The chip next to the title sums up the project. Below it you see how many episodes and recordings there are with their verdict counts, how many recordings are Not anonymized or Not segmented, and a Last 30 days line that counts uploads by result.
Data distribution charts the shape of your data: episode duration, episodes per recording, frame rate, cameras, subtasks, task frequency, outcomes and hours by collector or scene. Many bars are clickable and open the matching episodes. A long tail of slightly different task phrasings, for example, tells you the AI proposals need a review pass.
Work through the issues
Issue types at the bottom lists every check that warns or fails, with a short explanation and the number of episodes it affects. Warnings are real findings, not noise to hide. In the screenshot below, a human recording shows five of them: possible faces left after blur, soft frames, underexposure, hands in view and idle time.
Choose a row to open Episodes filtered by that check. One thing to know today: this link lists only the episodes where the check failed, so for a check that only produced warnings, like these five, it shows no match. To see the affected episodes, open Episodes and set Verdict to Warning, as in the screenshot below.
Open one with Annotate to watch it in the episode workspace, and use Copy time to share the exact moment with a teammate. For a possible face, open the recording and look through it: the explanation already notes that most hits are usually not faces.
The ⋯ menu at the end of a row lets you Reprocess the affected recordings, Ignore for this project with a reason, or Assign the issue to a member. A Blocking chip marks a check your project treats as blocking. If recordings stopped during processing, a separate card lists them; choose Show failed recordings to see which step failed and why.
How quality decides what goes into a version
When you create a dataset version, episodes with warnings are included, and the dialog tells you how many. Episodes that failed a check and episodes whose checks have not finished are left out by default; you can let them in with Include episodes that failed quality checks or Include episodes whose quality checks have not finished. Human recordings must also have passed anonymization before their episodes can go in.