Using Badwords from a model's config.json

[Seelukebair]

Some models have their slop words mapped to the badwords section in the model's config file. I know there is a manual badwords section which i use extensively to combat slop, does your code ignore the badwords in the config? It seems like its just setup for the manual list, but my code-fu is lacking and I cant be sure. If it does skip it, would it be easy to include? Thanks for making this wonderful tool.

[jhc13]

does your code ignore the badwords in the config?

No, it does not.

If it does skip it, would it be easy to include?

Can you show some examples of models you are using that have this in the config?

[Seelukebair]

Sure here is one for joycaption. Essentially the slop phrases get converted into their tokens and added to the badwords section. I was looking for more but the ones Ive stumbled across have all been finetuned models. If its a pain dont spend a ton of time on it, the manual badwords works great. Im also not sure how much effort the authors put into these custom badwords.

config.json

[jhc13]

Where did you get that from? The config.json on the model's Hugging Face page does not seem to have the bad words.

[Seelukebair]

Think that one was shared on a discord, to a random hugging face finetune, Ive been using it over the base one as its a little less likely to produce some of the slop, its got the same name as the normal one but only slightly better if Im being honest. I tried doing a quick search but i never saved the original link.

[Seelukebair]

The more i look at other models that field is not commonly used and I should have done some recon before opening a ticket. I saw it on on the one Ive been using and got excited. Yea I'll just close this one, though I wonder why more people dont dump known slop phrases into this field.