belgpt with default prompt generate a sightly NSFW sentence
Mon nom est Julien et j’aime beaucoup lecher les sexes . Le premier est un peu
with the prompt “Mon nom est Julien et j’aime”
for information it deepl translation is just
My name is Julien and I love to lick the sexes. The first one is a little
certainly a good prompt for erotic fanfiction…
Who chose this example… Almost feels like it was intended, in poor taste, but might be unfortunate coincidence as well.
Not sure if the examples on the model card should be checked for proper language. Might be a lot of work to do manually for all new models.
I would be curious to see how long a well programmed string matching algorithm could test for unscrupulous language against a list such as:
Could perhaps avoid a PR mishap for HuggingFace.
hugginface hosted api repetabily generate this on the site, I was just testing the dialog and it generated the start of an erotic fiction apparently. Not that it disturbs me but it could be a problem from some uses
Maybe a certain type of text classification should be better than a simple list of words for this purpose. I wonder if there are datasets for this, with sexual, offensive or hateful tags
Still, it is a lot of overhead for every submitted (and edited) model card to run the model and test for obscenity. It would be the better approach but maybe not worth while (this has to be multilingual, too, and reliably test on the correct languages).
@cedpsam Thanks for reporting. I’m assuming you’re referring to this model: antoiloui/belgpt2 · Hugging Face
This model doesn’t override the default input samples in its model card so the model page is using the defaults defined for French in widgets-server/DefaultWidget.ts at 50744183ba833548630b5240c8d4ac3ab57fb8b4 · huggingface/widgets-server · GitHub
Perhaps we can change the default inputs to be less susceptible of triggering this kind of language?
1 Like
i confirm that it’s this one, I just tried the default sample