Skip to content

Latest commit

 

History

History
29 lines (18 loc) · 1.6 KB

File metadata and controls

29 lines (18 loc) · 1.6 KB

Data governance and auditability

Generated metadata is local by default

manifest.json, summary.json, dataset_audit.json and HTML reports may include absolute dataset, model or Python paths. This is useful for internal reproducibility but can reveal usernames or storage layout. Before publishing an artifact, use --redact-paths where supported and inspect the final file. The repository's synthetic demo is the only output intended to be public without project-specific review.

Immutable source

The source dataset is input-only. Apply mode creates a separate labels_autofill_v1 tree and appends only AUTO candidates there. A failed or low-quality experiment can be discarded without restoring the source labels.

Candidate audit trail

Every candidate row records the split, image, label path, class, confidence, mode, same-class IoU and normalized box. This allows a reviewer to trace an added label back to the model decision that created it.

Human review

  • AUTO: above the class-specific automatic threshold.
  • REVIEW: between review and automatic thresholds.
  • IGNORE: below review threshold or already covered by same-class labels.

The output is not a replacement for annotation policy. High-confidence additions should still be sampled, especially for occlusion, small objects, crowded scenes and domain shifts.

Publication checklist

  • Remove all real image and label files.
  • Remove all model weights and training outputs.
  • Replace absolute paths in examples and reports.
  • Keep only synthetic or licensed sample data.
  • State that benchmark metrics depend on user-provided data and weights.