Manage dataset versions
Create stable dataset membership snapshots, inspect earlier versions, and choose the version an offline test runs.
Dataset versions let you change a regression set without rewriting earlier test inputs. You will create a new membership snapshot, inspect the prior one, and select the version used by the next test run.
Prerequisites
- A non-legacy dataset with a defined schema
- At least one example or production trace you can add, or one current example you can remove
How versioning works
The platform creates versions automatically. Adding a batch of examples or removing a batch from the latest snapshot creates one new version. There is no separate Create version action.
A version records which examples belong to the dataset at that point. The schema is one locked contract shared by every version. Earlier versions remain read-only and keep rows later removed from the latest version.
Create and inspect a version
Note the current version
Open Datasets, select the dataset, and read the version selector in the
page header. Record the current vN and its example count.
Change the latest membership
Add one or more traces through Add production traces to a dataset, or select rows on the latest dataset version and choose Remove.
Keep related additions or removals in one batch when they should form one reviewable snapshot. Separate batches create separate versions.
Compare the snapshots
Return to the dataset and open the version selector. Confirm that a new version appears with the updated example count. Select the previous version and verify that its membership has not changed.
Removal actions are available only on the latest version. Switch back to the latest snapshot before making another membership change.
Select a version for a test run
Open the dataset’s Tests tab, choose Run Test, and select a test config. In Start a New Run, choose the intended Dataset Version.
Leaving the selector at its default resolves the latest version when the run starts. Choosing an explicit version is safer when teammates may edit the dataset at the same time.
Verify the pin
After the run starts, open Tests. The run row shows the resolved dataset version, and the run keeps that exact version even if the dataset later changes. The same row also records the exact judge versions resolved for the run.
Troubleshooting
- If a run says the dataset version has no items, choose a version with at least one example.
- If the version selector or editable schema is missing, confirm that the dataset is not a read-only legacy dataset.
- If a removed row still appears, check the selected version. Historical snapshots intentionally retain rows that were live at that point.
- If an addition produced more versions than expected, it was submitted in multiple batches.
Next step
Run an offline test with the version you just verified.