SLOP FIGHTER · published
This improvement ties one thread I've been unsure was able to be tied. Now, animal body parts and biological traits are better contextualised for the LLM's understanding. This means they can, and will, narrate the way they use themselves more effectively. A monster with camouflage, like a chameleon, will reference it when an attack against them misses, for example.
Underlying the narration are extensive semantic graphs, which are simply blocks of words sorted into categories, like body parts, movement styles, and mutation vocabulary. The LoRA adapters are trained on sentences constructed out of these graphs. Each word, in consideration of the sentence structure it was trained on, is given numerical weight, and at runtime the LLM relies on its internal mechanics to decide which word to generate.
Its decision-making when it comes to generating words can be influenced, kind of guided, by careful prompting. Essentially, if you give the LLM enough of the information it needs in the right way it can start to build the required context to generate narration that reflects its character.
It's a problem I'm familiar with as an author. How do I get the reader to generate an image in their head based on the words I am using? How do I make this person understand? It's kind of insane to think about how the challenges are the same for what is just a machine that generates words.