Drop the dataset
YOLO folder with data.yaml, COCO JSON with images, VOC XML with images, or class folders.
Datasets
Dataset Validator
Training jobs fail hours in because of one malformed label file. This validator reads a detection or classification dataset and lists every problem with the file it is in: images without labels, labels without images, class ids outside the class list, boxes outside the image, zero-area boxes, duplicate boxes, unreadable images and mismatched extensions.
Loading the tool…
How it works
Read the report
Problems are grouped by type with counts and file names; click one to see the box drawn on the image.
Fix or export
Download a cleaned copy with the chosen fixes applied (drop bad boxes, clamp to image bounds, remove orphan labels) and the report as JSON.
Formats, limits and the numbers that matter
Checks
Missing pairs, empty label files, class id range, coordinate range, zero or negative size, duplicates (IoU above 0.95 with the same class), image decode failure, EXIF rotation that changes width and height.
EXIF rotation
Phone photos often carry an orientation tag; a loader that honours it swaps width and height and every box is wrong. The validator flags those images so you can apply the rotation first.
Speed
Ten thousand label files check in a few seconds; image decoding is the slow part and runs in the background.
Common questions
Should empty label files be deleted?
No; an empty file is a valid 'no objects' image and helps against false positives. The validator distinguishes empty from missing.
What about COCO iscrowd?
Crowd annotations are reported separately and kept.
Can it validate a JSONL chat dataset?
Use the JSONL Inspector or the LLM Dataset Builder, which validate message structure.