Free tool
Dataset Splitter
Split a COCO JSON file into train, validation, and test sets with stratified sampling. Everything runs in your browser.
Split dataset
Upload a COCO JSON file and split it into train, validation, and test sets.
Frequently asked questions
- What ratio should I use for train, validation, and test sets?
- 70/20/10 and 80/10/10 are common starting points. Use a larger test set when you have few images per class, so each class still appears in the evaluation split.
- What does stratified splitting do?
- Stratified splitting keeps the class mix of each split close to the full dataset. Without it, a rare class can end up missing from the validation or test set.
- Can I reproduce the same split later?
- Yes. Set a random seed. The same file, ratio, and seed always produce the same train, validation, and test sets.
- Are my files uploaded?
- No. The splitter only reads COCO JSON files of up to 50 MB, and the split runs entirely in your browser. Nothing is sent to LabelOp servers.