getting a language model to write lipograms by simply zeroing out the probability of any token in the vocabulary that has a particular letter in it (in this case, 'E')

· · Web · 1 · 3 · 17

doesn't do so well at the inverse task, i.e., generating with the probabilities of any token containing a vowel letter OTHER than 'E' zeroed out

@aparrish have you seen @GLOSSATORY? it also does the former and sometimes (randomly) the inverse on initial letters

@ranjit @GLOSSATORY I have, yep! I imagine it works along similar principles

@aparrish @ranjit yes, both it and @gravidum_cor do this, but at the letter level rather than word, so they generate a few neologisms. also has a couple of other sources - a model trained on inputs without “e”, which generates babble, and posts from the main which are accidentally compliant.

Sign in to participate in the conversation
Friend Camp

Hometown is adapted from Mastodon, a decentralized social network with no ads, no corporate surveillance, and ethical design.