Skip to content

dataset_validate

Deep integrity checks for YOLO and COCO datasets.

Both formats are translated into the shared internal DatasetInventory model before validation. This keeps readable-image, bounding-box, class, orphan, and metadata findings aligned across formats without changing the public tool schema.

Inputs

  • dataset_path
  • format: auto, yolo, or coco (voc remains unsupported)
  • check_image_readable
  • check_bbox_bounds
  • check_class_id_range
  • strict

Outputs

  • valid
  • structured errors and warnings
  • one-line summary
  • optional remediation_script

Bounding boxes are checked in normalized image-relative coordinates. YOLO class IDs are checked against nc; COCO category IDs are checked against the declared category registry. Missing or corrupt COCO image files are reported with the same Image unreadable semantics used for YOLO.

Remediation remains explain-first. The optional clipping script is emitted only for YOLO text label files; COCO JSON findings are reported without generating an unsafe format-specific rewrite.

Examples

{"dataset_path":"~/data/mini_yolo"}
{"dataset_path":"~/exports/coco128","format":"coco"}
{"dataset_path":"~/data/warehouse","strict":true}
{"dataset_path":"~/data/retail","check_image_readable":false}

Common Errors

  • unsupported format other than YOLO or COCO
  • missing dataset path
  • malformed format metadata

dataset_inspect, dataset_convert