I have to say this is working significantly better than openrouter. I wonder if my internet could have been causing issues. the one thing I'm trying to get to work now is getting the kobold entities to speak in third person. I have it saying, kobolds always speak in third person in the entities description and I have it as entities that speak in third person always start with their name first when speaking. with your prompts the game is still kind of pulling me along but I have much more control now. I think I just need to try and fine tune it a bit.
Glad to hear that it works better with the local host! I doubt it was the internet speed though, keep in mind that a lot of people use those models. Even the paid ones have their fair share of work.
If you are using LM Studio at its' default settings, it will automatically throw out the excessive memory once it nears the limit. The default setting for that is that the AI throws out the middle part, keeping the start (with the world rules) and the more recent prompts in memory.
For the weird sentence endings, you can try adding something like this to your prompt maybe? I got this from ChatGPT and it sounds like it should work.
> Only describe events, actions, and dialogue that are relevant to the player's current situation. Maintain an open-ended format.
As for the kobolds, something like this in the AI section of their entity entry should do the trick:
> The kobolds always address everybody in third person. Even when they talk about another of their kind, they use third person only.
they have made it much better! I also added Do not choose or suggest the player's next action. For example, if the player chooses to lay down do not respond that they fall asleep. Sadly, I had to switch from local host as I could not for the life of me get to enact even slightly sexual scenes. However, using your prompts and adding a few of my own I got microsoft/wizardlm-2-8x22b to work really well through openrouter! Now to hope it stays working well lol. Even with it not being an instruct. it seems to handle the stats just fine for the moment.
To be honest, I think that the model no longer needs to be an "Instruct" model necessarily. The Instruct part was a necessity back before 1.1.0 because the game used a JSON format to control everything, but now that the game can just use plain text, any model should do the trick. The most complicated one would be for the stats, and even that is doable. I'll try out using a regular model for the game and maybe update the guide so that any generic text model can be used.