Documentation for filtering illegal content
Hello.
I read the paper section indicating there was NSFW filtering. But could you clarify which code you used to filter out illegal content from this dataset? There is a formal investigation underway for LAION-5b, its filtering code is public and very basic, and obviously could not catch all illegal content.
Could you clarify which code you used to filter DataComp?
If it's code of similar quality than LAION's, I suggest you immediately TURN ON GATING for the dataset and put huge disclaimers about the illegal content in the Dataset Card.
As you mention in the paper, if this is intended for research in filtering illegal content, could you clarify who has the right to download illegal content to do this research?
Thanks!
Revisiting this given the recent (expected) news about LAION 5B. It appears to me that the dataset creators (incl. HuggingFace) do not take this matter seriously.