2 of 7 · Data workflows

Datasets and annotations

Create datasets, ingest samples, import annotations, verify readiness, and preserve split integrity.

Build a training-ready dataset#

Inspect capabilities
Confirm the supported data type, task, annotation format, and organization limits before transferring a large dataset.
Create the dataset
POST the required name, description, and format to /datasets/datasets/. Keep the returned dataset UUID as the stable identifier.
Upload or register samples
Use the endpoint appropriate to uploaded files, imports, or trusted local MCP paths. Stable filenames are essential for annotation matching.
Import annotations
Use the production COCO, mask, CSV, or platform JSON workflow that matches the source data. Review unmatched and invalid counts.
Prepare and verify
Check sample processing, labels, annotation coverage, and training/validation/test counts before starting compute.

Choose one selector
Send either sample_ids for an explicit bounded list or sample_query for a server-side split, search, and filter selection. Never send both; omit both only when the full dataset is intended.
Choose split behavior
Keep preserve_splits true to copy the selected samples’ training, validation, and test memberships, including overlaps. Set it to false when the clone should start with no split assignments.
Verify the clone summary
The server evaluates sample_query against the current source when you submit. Save the returned private dataset id and treat clone_summary.sample_count, split_counts, and preserve_splits as authoritative before using the clone in a pipeline.

Preserve subject or patient grouping when leakage would inflate evaluation results.
Use append-only rebalance when new samples should be assigned without rewriting established split membership.
Use split locks and recorded seeds where the workflow supports them.
Export a source snapshot before a bulk annotation update.