Score Studio
Get started

Manage dataset assets

Curate images and video frames inside a specific dataset version. Score keeps reversible cleanup separate from permanent removal, so routine quality work does not silently erase annotations or provenance.

By the end of this guide

You can locate weak inputs, add new evidence, archive or restore assets in bulk, permanently remove data when required, and preserve annotation text when moving a VLM dataset between systems.

Core workflow

  1. Open a version

    Choose Data → Datasets, open a dataset, then select the version whose asset contract you want to inspect.

  2. Find the evidence

    Search by path or filename, filter labeled or unlabeled assets, and narrow the inventory by train, validation, or test split.

  3. Curate safely

    Archive questionable inputs first. Archived assets leave the active inventory but remain recoverable with their engineering context.

  4. Continue the loop

    Open Annotation studio to annotate or review the selected version, then create a new immutable version before training or evaluation.

Choose the right action

  1. Add assets

    Upload files or connect external storage to extend the current version's source inventory.

  2. Filter and select

    Combine filename search, annotation state, and split filters before applying a bulk operation. In Annotation studio, select unwanted video frames directly from the image strip.

  3. Archive

    Use for duplicates, weak samples, test uploads, or uncertain data. The asset can be restored.

  4. Restore

    Return an archived asset to the active inventory with its retained context.

  5. Delete permanently

    Delete selected images or video frames together after one confirmation. Their files and attached annotations leave this version.

Annotations and VLM text move together

Annotation import and export are version-scoped. Detection and segmentation datasets retain their boxes, polygons, masks, classes, and confidence metadata. Phrase grounding, visual question answering, and captioning datasets also retain the text, answer, or caption spans linked to each image and region. Review the import summary before accepting a schema mapping, then export a backup before destructive cleanup.

Specialist vision

Images or frames · classes · boxes · polygons · masks · split membership

Multimodal data · external training

Images or frames · prompts · answers · captions · linked regions. Check the chosen export format. These schemas do not provide an in-app VLM trainer.

Expert section

Identity, lineage, and destructive operations

Asset operations are part of the dataset contract. Automation should address assets by immutable ID, preserve version lineage, and treat archive and permanent deletion as different state transitions rather than interchangeable cleanup actions.

Quality and operating signals

  • Duplicate, corrupt, and unmatched asset counts
  • Annotation coverage before and after curation
  • Split membership and downstream version references
  • Import rejection and schema-mapping rate

Common failure modes

  • Matching assets only by a non-unique filename
  • Permanently deleting evidence referenced by a release
  • Losing linked VLM text during format conversion
  • Mutating a split in place and introducing benchmark leakage

Expert operation checklist

  • The operation is scoped to the intended dataset version.
  • Selected assets were resolved by stable identity.
  • Referenced annotations and multimodal text were exported if needed.
  • Permanent deletion has an explicit owner and audit reason.

Deletion contract

Archive first. Delete only after confirming the asset and its annotations should leave the version permanently.

Archive is the default cleanup action. Permanent deletion removes the selected asset and its associated annotations from that dataset version and cannot be undone from the interface. Export a backup first when retention, reproducibility, or audit evidence matters.
Manage dataset assets · Score Studio