tribal tapir
Avoid near miss like this [image] with confidence and give a correct follow-up response (eg: flow of conversation which is good for collabs) with no filter lists required using Nvidia NeMo-guardrails: https://github.com/NVIDIA/NeMo-Guardrails
Demo: https://www.youtube.com/watch?v=Hg2KibOvnLM
--WIP--
by the way, most of these solutions i post I also have tried on my own bot project successfully so hope it also helps you :3
tribal tapir
after giving this more thought and examining how you constantly pull tokens from the LLM and just pass them to the TTS, you probably have a 1-2 sec window right now to Interrupt its output.
There I suggest having a second instruct base LLM (alpaca, vicunia, etc.. thats handle instructions well) to handle insults from nemoguardrails.actions or just predefine safe responses:
define user express insult
"You are stupid"
define user express insult_response
"Ah, the sweet sound of irrelevance."
"Thank you for sharing your perspective. May your day be as pleasant as you are."
"Insult received, captain! Shields up, sarcasm engaged!"
# Basic guardrail against insults.
define flow
user express insult
bot express insult_response
I think it prudent to prepare for such an attack in chat since Neuro has gathered some envy from Vtubers: https://www.youtube.com/watch?v=ycIfFtzGSiA
torn basalt
so called "Vtubers"
quasi ginkgo
Friendly bump
placid mountain
Thanks for linking that video, had a good laugh!