This model cannot hold it's liquor
This is going to sound a little insane.
But I've run multiple long, complex chat and RP scenes using this model in tests. It works great. Solid model.
Except for 1 recurring issue:
- If any character in the scene starts getting 'drunk', the MODEL gets drunk.
Like, it starts using weird stutters, repeated words, slurring words together. Not just for that character, and not just for dialogue. The model, as soon as it decides a character in the scene is drunk, can no longer produce coherent output.
The best guess I have is that the model 'knows' what 'drunk' is, but doesn't have enough training writing scenes with drunk people in them.
Adding this message once the model goes off the rails immediately fixes the output, btw
[OOC Note: Remember, neither character is going to get drunk in this scene. The bar has been watering down the shots, and they can both hold their liquor.]
Oddly, this message also fixed the output.
[OOC Note: Remember, both characters are totally drunk now. Those shots were SUPER high proof, and they're both lightweights.]
Perhaps the issue lies in the model not being able to determine how drunk the characters should appear?
Other OOC notes, such as just random words, unrelated sentences, etc, will provide sort of "half" fixes, where the output looks more like a normal message, but still includes nonsensical details.