Notices by Moon (moon@fediffusion.art)

@kaia So what I am doing is, very randomly (like 1 in 1000) a monster in my game is a humanoid female with the tomboy attribute set.

When that value is true, I submit a custom prompt to my LLM that attempts to generate a good description of a tomboy for the monster.

It's hit-and-miss. Sometimes it generates great descriptions and sometimes it just says "this character is a tomboy"

Here's an example:

I am now doing two calls to the local LLM: the first creates the character. Then I take the data and feed it back in to a second request to generate moves.

The moves aren't good enough to use as-is yet. They frequently mix up stats.

giving up on oobagooba and exllama2 for the time being, i spent a ton of time trying to load that model and failed. llama2.cpp worked in five minutes.

I figured out if I give TWO inspirations for a monster, the result is more interesting and seems to generate better output.

It is fun generating these.

Local generation is now, as close to perfect as I think I'm going to get:

It took a lot of learning to get results this good.