Dataset import and Reprocess (`/p/<project>/datasets`)

Bring an already-labeled dataset into the current project and re-run the backend's machine proposals without overwriting human work. This page covers the screens; formats, class mapping rules and the API/CLI path are in the Dataset import guide and the Reprocess guide. These pages exist only when the backend supports dataset import; otherwise there is no link to them and nothing is requested. Compare with Ingest, which brings in unlabeled images.

Formats

The backend decides which formats it can read and serves the list. A typical backend serves Detect automatically, YOLO (a data.yaml layout), COCO JSON and an OpenProcessor export (the dataset the app's own Export page writes). The wizard only offers what the backend serves.

The import wizard (`/datasets/import`)

Source. Type a folder path the backend can read (the allowed roots are listed under the field, and the preview tells you if the path isn't usable), or upload an archive. An archive larger than the smaller of the backend's own limit and the deployment's OP_CROPWRIGHT_DATASET_UPLOAD_MAX_MB is refused before it is sent. Pick a format, or leave Detect automatically.
Preview. A moment after any change, the backend previews exactly what you've chosen: splits, totals, an estimate, facts about an OpenProcessor export (such as whether a test split is present), and a list of issues grouped by severity with samples (file and line). If an issue blocks the import, Start is disabled; a Start anyway checkbox appears only when the backend says forcing is allowed for those issues.
Class mapping. One row per class in the dataset, matched to your project's classes by name only. Each row shows the dataset class, its box and image counts, the backend's suggestion and how it matched, and the target the current choices resolve to. For each row choose to map it to an existing class, create a new class, treat it as a region, or skip it. A row that has boxes but no target yet is highlighted. Use suggestion copies the suggestion into a row; Accept suggestions lets the backend fill every row it can; and Use the mapping from a previous import copies an earlier completed import's choices by class name. Two dataset classes merge by picking the same target class on both rows.
Options. What to do with each image (import as-is, or also run the backend's detectors and VLM, with each mode's description shown), how far to trust the imported labels, how to treat region parents, whether to freeze the dataset's test split as the project's holdout, and some advanced settings. An option you leave alone is not sent, so the backend's own default applies.
Start. A confirmation shows the previewed totals and the request is tied to that preview. Refusals use the backend's own words: an already-imported dataset links to the earlier job, an interrupted one offers Resume, a dataset that changed on disk asks you to preview again, and an incomplete mapping highlights the rows still missing.

To import into a brand-new project, create the project first on /projects; the wizard always imports into the project you're in.

Imports and the job view

/datasets/imports lists past imports with their status and a New import button. An import's page (/datasets/imports/<id>) shows:

the status, a progress bar with throughput and ETA, and whatever the job is waiting for;
the report counts, the issues and the per-image entries (paged, filterable);
the class mapping the job actually used;
the error, if the job failed;
Next steps the backend suggests, each behind a confirmation;
Review imported labels, which opens the Imported tab.

The page follows the job on the backend's own polling cadence and wakes immediately when the backend announces progress. Cancel, Resume and Undo each ask for confirmation, and are offered only when the backend says they are allowed; when one is not, the backend's reason is listed under the buttons. Undo does a dry run first and shows what would change, including that your own edits are kept, then lets you choose whether to also remove the imported images and deprecate the classes the import created.

Reprocess

Reprocess re-runs machine proposals on items that are already in the project. It respects the lock rule: anything a human has labeled or edited is locked and is skipped, not overwritten, and the result counts how many items were skipped for that reason.

One item. In an item's Details panel, choose Reprocess, tick the scopes to re-run (detect, region, VLM, embed; for regions, redetect or reverify), and confirm. The updated item replaces what you were looking at.
One image. On a card's expanded view and on a region card, Reprocess image… re-runs the proposals for every item cut from that image.
A selection. On /clusters/<id>, select crops and use Reprocess in the toolbar. The backend does a dry run first, you see the per-scope counts (selected, locked and skipped, queued, failed, not found), and then confirm to apply. A large request becomes a job you can follow and cancel.
After a profile change. Activating a region profile can suggest a reprocess; see Region profiles.

Lock badges

A lock icon marks a label (on a crop card) or a box (on a region card or in the box editor) that a human has set and the backend will not overwrite. Hover it to read the backend's lock rule (what makes a label locked). The Reprocess scope names come from the backend as well.

Import provenance

An item's Details panel shows how an imported item got here: whether its label is locked, the dataset split, which import(s) brought it (linking to the import), when, what the import proposed, and whether it sits on a negative frame. A row appears only when it has a value.