Multiple datasets, copying and combined export
A project can contain several independent datasets that share its task, labels and attributes. Each dataset keeps its own media, split, versions and history. Copy creates an independent working set while reusing media bytes; Export all datasets as one merges selected image datasets under the shared schema and an explicit split strategy.
What is shared and what is separate
| Shared by the project | Separate per dataset |
|---|---|
| Task type, labels, prompts, attributes and keypoint template | Media, annotations, train/validation/test split, versions and history |
Use the plus control beside the dataset tabs to create another set. The first six datasets appear as tabs; additional ones move into a picker. Phones show one dataset at a time. Rename, Import, Export, Copy and Delete live in the active dataset’s menu; Delete is offered only when another dataset will remain.
Copy before a risky change
Copy dataset can include all, annotated or unannotated media; images, videos or both; and optionally their annotations. The copy is independent after completion. Media bytes are content-addressed, so identical files are shared in storage. Cancellation rolls the operation back instead of leaving a partial dataset.
Export all datasets as one
From the Datasets list, open Export all datasets as one, select the source datasets and choose a compatible image format. Same-named files from different datasets receive collision-safe export names. All selected datasets already use the project label schema, so classes reconcile without a label-mapping step.
Combined split strategies
| Strategy | Result |
|---|---|
| Preserve | Concatenates the existing train, validation and test subsets; unsplit images are placed in train with a warning |
| Recombine | Ignores existing assignments and computes one fresh split across the selected images |
| Flat | Writes one dataset with no train/validation/test folders |
Recombine groups byte-identical images into the same subset, preventing a duplicate from crossing between train and test. The archive includes a combined manifest that records the source dataset and media id for every output image, the split strategy and any warnings.
Combined export is deliberately image-only
Native videos and Pending AI Review drafts are excluded. Use a single dataset’s Supervisely Video, MOT, MOTS or Datumaro export for video data. Use Export project when the goal is a lossless AnnotateIt transfer containing every dataset, video, frame annotation, track and project setting.