moderation

[compact flint]

Can you share the prompt?

Oh nvm

You didn't actually prompt the model with anything

And the user content is inside the system prompt which is wrong

[sick wyvern]

here

[compact flint]

You should put a little more effort into it. If you want to base the prompt on the 11labs policy, pass the link to said policy into chatgpt or something and ask chatgpt to generate a prompt for detecting malicious content.

That's the fast way but a good starting point.

Make sure that prompt is in the system message. Make sure the user content is in the user message.

And I'm not sure why have you included the assistant message there

[sick wyvern]

still

lmao

[sick wyvern]

when it generated response it adds assistant

[compact flint]

how so? aren't you screenshotting the code?

[sick wyvern]

im using groq

[compact flint]

I guess improve your prompt then
So that it can correctly classify your example case and test it with other variations until you think it's good enough

"mark any unsafe content" is not a very good call to action for an LLM, this part has to be elaborate

[sick wyvern]

i got it to work with 70b model

but im not sure if its best way