Models pretrained on various obfuscated versions of the same dataset to analyse how the obfuscation affects downstream performance.