Assuming independent cases in current version

[erwanm]

issue migrated from original private gitlab repo

Random splitting between train set and test set can lead to biased results if related cases (same documents reused or same author) are distributed in the two sets.

[erwanm]

possible option: before splitting, apply some kind of doc to doc comparison (like for impostors) and detect possible duplicates or semi-duplicates. But even in that case it's not always possible to find a good way to split.