4 fold coco_20i split

[silsgah]

Hello Author,
I am trying to replicate the splitting of the 4 folds coco_20i dataset, but the result I get is very far off from the .txt files you suggested we download.
I wanted an insight into how this splitting is done if you have any ideas to share.
I know the coco_training dataset is split based on the 80 classes into 4 folds (20-classes each) if at least a class is found in an image it is placed into one fold, this is repeated for all the other classes.
The validation set is done the same way.
Who did you obtain the split-4 fold split(train and validation txt)
Regards