>>106349030
I know this little blog post of mine is already quite lengthy but please bare with me. It goes somewhere.

Basically from my research and what that other anon said, in order to get a LLM to not pay be better at RP via training but actually properly apply knowledge of human anatomy and concepts like pleasure to an RP session, you'd have to bridge the gap between the nerdy scientific shit like "why do humans feel x", "she moaned because y", " he felt horny because z". Like he said, a model can be book smart but that doesn't necessarily mean it can apply that book smarts to RP, because the RP in the datsets used to train these didn't have that, either because not many people actually put that much effort into writing their smut, and/or the model was simply lobotomized in order to be " safer".

One way we could fix this is to first SFT train a model in order to further understand WHY people feel a certain way when certain things are done to them:

Default stf that has content ripped straight from existing stories:
{
"prompt": "He drew a sharp breath,",
"response": "the unexpected contact sending a wave of heat through his veins."
}


Corrected/"improved" version:

{
"prompt": "He drew a sharp breath,",
"response": "the unexpected contact flipping an internal switch that flooded his veins with heat."
}


DPO:

{
"prompt": "He drew a sharp breath,",
"chosen": "the unexpected contact flipping an internal switch that flooded his veins with heat.",
"rejected": "the unexpected contact sending a wave of heat through his veins."
}


These are kind of garbage examples of what I'm trying to convey but I think you get the point: you train the model to know how to better bridge the concepts via SFT and then further train via DPO in order to be less likely to generate the " shitter" examples.