Romain Beaumont
Romain Beaumont
reconstruct is too slow for memory mapped indices so it would have to be a filtering on url and text to begin with
partially implemented would be nice to make it more generic
at least under an option
if doing that, put the site referal like `"referer": "https://www.jf-studios.com/"` as headers to bypass protection such as in https://jf-studios.com/wp-content/uploads/thon/drawing-anime-lady-kawaiiiii-anime-girl-drawing-sketch-in-2019-pinterest-drawings-of-drawing-anime-lady-300x170.jpg (do not click, NSFW outside of the website) (not sure if...
outputting filtered parquet files would make a lot of sense
(and possibly webdatasets)
what specific weight do you mean ?
did you retrain a model with the new version ?
Do you mean https://github.com/robvanvolt/DALLE-models/tree/main/models/taming_transformer ? Let's fix it there then
could be adapted to do that, in particular these parameters https://github.com/chg-hou/EnMicroMsg.db-Password-Cracker/blob/master/password_cracker.c#L22