Ensure all data splits contain at least one sample

[juditacs]

This is important when using toy datasets. Valid and test splits can easily end up empty.

I'll probably change the sampling method to fixed sized samples instead of random sampling and using threshold.