![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
Currently LLMs are pretty formulaic about making up narratives on their own, and their style is obvious after one views enough of it.
Inspired by the effect in Stable Diffusion, I tried some negative prompts on my current favored 7B parameter LLM, and the result of all the negative prompts I came up with in order to make the stories less awful turned out to be unintentionally humorous.
The punch-line: the LLM still tried really hard to sneak in moralizing endings. I presume this was a function of the training corpus. GIGO in action.
Inspired by the effect in Stable Diffusion, I tried some negative prompts on my current favored 7B parameter LLM, and the result of all the negative prompts I came up with in order to make the stories less awful turned out to be unintentionally humorous.
Write a story. Avoid flowery words. Avoid an explanatory paragraph at the end. Avoid moral tales. Avoid a fairy tale ending. Avoid mention of journeys. Avoid mention of things lasting a lifetime. Avoid mention of indelible marks. Avoid meta explanation of the story. Avoid mention of unimaginable challenges. Avoid mention of testaments. Avoid mention of stories as reminders. Avoid mention of powerful lessons. Avoid mention of forever.
The punch-line: the LLM still tried really hard to sneak in moralizing endings. I presume this was a function of the training corpus. GIGO in action.
