Dataset administration

A bootstrap administrator creates datasets. Within a dataset, the data-admin role manages configuration and membership; annotator and reviewer are separate permissions. Any member can inspect images and view statistics.

Create or reuse a schema

Setup creates a fresh dataset. Copy schema from lists datasets where you have DataAdmin access and previews classes, workflows, and ordered keypoints. Creation also requires bootstrap-administrator authority. A missing, invalid, or inaccessible source fails creation without silently switching to an empty schema.

Copied	Omitted
Class and task IDs, names, relationships	Images, annotations, reviews, assignment history
Workflow settings and instructions text	Source users, roles, ingestion paths, balance settings
Skeleton specifications and ordered keypoints	Tutorial image references and prelabel bindings/models
	Migration configuration and guide links, original dataset identity

Copied skeleton workflows use ordinary annotation and retain their review settings. They do not require imported guide boxes or migration confirmation. The source dataset keeps its migration configuration and history. This applies to newly created copies; existing datasets are not changed retroactively.

The datasets are independent after creation. Matching IDs, unchanged definitions, and matching export selections produce compatible class indices and keypoint layouts. Later edits or different selections can break that compatibility.

Configure a dataset

Administration groups changes into Overview, People, Images, Schema, Automation, Snapshots, and Export. Configuration edits remain staged while switching sections. Resolve validation errors and save before starting work that requires saved configuration.

Section	Use
Overview	Dataset identity and configuration overview
People	Assign annotator, reviewer, and data-admin roles; bootstrap-admin protections still apply
Images	Configure relative filesystem roots, upload folders, run ingestion, inspect job results
Schema	Define classes and box/skeleton tasks, keypoint order, workflow settings, and instructions
Automation	Configure assignment balance, models, and dataset hint generation/removal
Snapshots	Create and download annotation snapshots
Export	Preflight, build, inspect, cancel, and download private detection/pose jobs

An approval workflow requires one reviewer. Without approval review, annotation submission completes the task. The legacy allowReviewerCorrections setting remains readable but no longer disables corrections. Instructions display title and text; configured example images are not rendered.

Assignment balance uses an absolute maxDifference between enabled tasks. Disabled tasks are excluded, and there is no independent class-level balance. See assignment for count and boundary rules.

Add images

Use relative filesystem image roots or upload a browser folder, then ingest. Ingestion indexes image bytes and deduplicates by BLAKE3 identity. Existing image IDs and known duplicate paths survive reconciliation. A filename alone is not image identity. Ingest job status is process-local and does not survive restart as a durable job.

Raw image ingestion does not import external labels. Use dataset import for a new dataset from supported YOLO or COCO ground truth.

Inspect and return work to review

Inspect is a read-only gallery for every dataset member. Search and filter by workflow, class, or status. The matching image list loads automatically in batches; thumbnails load as they become visible. The gallery distinguishes the total match count from the loaded list and labels previous results while a new filter loads. Previous/next navigation follows those filters. Gallery filters stay in memory and reset on reload.

A workflow filter includes unannotated images with Pending status, so selecting a workflow alone may leave the count unchanged. All workflows → Completed requires every configured workflow, including disabled ones, to be completed; datasets without workflows match no completed images. Selecting one workflow checks only its status. Other status filters under All workflows match any workflow. Class filters match active annotations.

Toggle workflow, status, and geometry overlays without claiming or editing an assignment. Boxes and skeletons use class colors and labels. A visible box supplies the label for its object group; hiding it restores labels on the visible keypoints.

Reviewers and data admins use Return to review in the Overlays panel or drawer to select completed approval workflows and enter a nonblank reason of at most 2000 UTF-8 bytes. Selection is independent of overlay visibility. All selected workflows must be enabled, completed, free of a live assignment, and have valid review targets. Disabled choices explain their exclusion through the adjacent info control. Imported ground truth has review disabled by default; enable approval review before returning its completed workflows. A conflict leaves the whole selection unchanged.

Successful return preserves geometry and historical decisions, clears the final outcomes, and starts fresh review rounds. Every current object and full-image check needs a new decision. A failed request keeps the reason and exact retry identity. Discard or success clears the draft while keeping the controls available. This action records audit history but currently sends no notification.

Statistics and scores

Statistics opens as a dataset-scoped modal without releasing current work. It reports five exclusive task states: Pending, In progress, Awaiting review, Needs correction, and Completed. Imported excluded coverage is outside the completion denominator until explicitly included.

The score podium and rankings lead the view, followed by daily activity, dataset counts, assignment balance, task/class breakdowns, and throughput. Period and history controls compare contributors. Empty or unavailable data is distinct from zero activity. Acceptance percentages accompany review counts. Scoring defines per-object rewards, daily tiers, focus bonuses, rejections, and the compressed displayed score.

Daily streaks require 20 annotation items or 30 review items per dataset per UTC day. Each completed Objects item and Overview item counts once per user, image and task. A whole skeleton is one item; partial saves, Skip and unchanged history reconfirmations add no credit. Legacy image/task submissions retain one unit, shared with any later Overview receipt to avoid duplicate credit. Flames and day counts appear beside contributors in the leaderboard and in the top bar. The top-bar flame opens Statistics. Streaks are derived from contributor history; reduced-motion settings suppress the goal animation.

Snapshots and exports

Snapshots contain metadata and replayable annotation/audit history, including committed import provenance and scoring focus history. They omit images, authentication, user keybindings, and private import/export jobs. There is no native snapshot restore. Use the full-root backup procedure for recovery.

Dataset export produces verified archives with original images and complete labels. It has explicit completeness and round-trip boundaries and requires current DataAdmin access, including when downloading retained jobs.