Clean labels
Find naming drift, empty values, duplicates, and imbalance.
Local-only analysis · Dataset stays on your device
Is your dataset ready for training?
Labelyst checks labels, geometry, image quality, and split health locally, then gives you a repair plan and reproducible quality report.
Check a dataset
Review labels, geometry, image quality, and split health before training.
Processed on this deviceUp to 500MBZIP contents use this limit after decompression
What you get
Everything needed before training starts
Trust geometry
Validate boxes, polygons, masks, keypoints, and overlaps.
Inspect images
Flag blur, exposure, corruption, rotation, and duplicates.
Ship evidence
Export repair lists, reports, policies, and version comparisons.
Where Labelyst fits
Between annotation and training
- 1
Export
From your tool or vendor.
- 2
Check
Run Labelyst locally.
- 3
Review
Inspect findings and repairs.
- 4
Accept
Return or approve with evidence.
- 5
Gate
Reuse the same rules in CI.
- 6
Train
Start with a known dataset state.
Who it is for
A local acceptance gate for data quality
ML and data teams
Check a dataset before training consumes compute.
Annotation buyers
Accept or return vendor deliveries with evidence.
Engineering teams
Run the same quality contract in CI before training or publishing.
What Labelyst is not
No accountNo dataset uploadsNo annotation editorNo model training
Evidence from public datasets
Real, dated local CLI runs on named public inputs.
99/100
Raccoon object detection
217 annotations · Ready with warnings
90/100
Ultralytics COCO8 detection
30 annotations · Ready with warnings
70/100
DataHub CO2 fossil global
275 rows · Blocked
Results show the current check rules, not a third-party certification. Read the evidence.