Banned tokens list.
Hello, me again. Having another issue kinda, though I doubt it's one you'll be able to truly fix since the llm is starting to apparently use other languages. The llm is getting around the banned tokens in the "eye" series by using Spanish, or "ojos". I'll try banning ojos, but I just thought i should let you know regardless. I honestly wouldn't be surprised if it tries using other Spanish words, or other languages at this point...
Yeaaah, that's a classic one.
Some models get around it my making typos too, like his eyeglinted, etc.
the probability for them is too high, the model is SURE that what comes next.
The way to solve this is banning even more backwards, you will have to make a ban for each variation of eye and eyes for his, her and their to swerve the model away from writing that, and add a space to give the model the chance to just end the sentence instead.
Kobold devs said they were looking to an alternative to regex to have wildcards in bans, it will be a blessing for cases like this
Since they raised the limit on bans, I think it's safe to add these variations.
I will upload a test list, try it, I don't want to push it to the main one without making sure it works.
https://huggingface.co./Sukino/SillyTavern-Settings-and-Presets/blob/main/Banned%20Tokens%20%5BAnti%20Eyes%20Test%5D.txt
Oh, and maybe the model will try to add colors to the eyes next to get around this, already happened to me. If it comes to this, there is no solution until we have wildcards, I won't ban every color. LUL
I'll try out the test list and probably report back then.