I find the DavidAU (very prolific ‘tuner’ of LLM) stuff interesting, and am distraught that the model that produces amazing output for my flavor of game isn’t an instruct model so it just barfs out a nicely paced tale and then creates another complete narrative upon feeding it a prompt. I’ve put it out of my mind for the moment until I figure out how to train it to respond to instruction. I think it’s Dark Forest or Dark Universe flavor.
NEW FAVORITE: deepseek/deepseek-chat-v3-0324 on open router. It’s a paid model, I’ve run $2 to $4 an evening running a few hundred prompts through it as I tune my scenarios. I have it set to 100k/10k context/output and it very consistently makes it 40 or 50 prompts deep into scenarios while staying consistent with the scenarios outline and carrying early improvisations deep into the ongoing narrative without too many “LLM: Please remember x,y,z” reminders. It tends to like to break its possible choices section into single lines, but that is very likely my prompt customization coming back to bite me. It”s doing better after I specifically told it to avoid line breaks while presenting its, often elaborate, choices. Deep into a scenario it will start presenting choices as if the scenario had just launched, but a custom prompt, even a simple “continue scene” will often get it back on track for awhile. It’s can be erotic without making the scenario participants immediately rush to humping, though it does tend to shift my “arousal” settings up to the max within three or four prompts due to all sorts of reasons (temperature plays a factor in one scenario and sweating has now joined adams apples and forearm hair as one of the most erotic things on the planet.10 out of 10 LLM’s concur)
I’m currently beating a 12.??gb version of DavidAU/LLama-3.1-128k-Darkest-Planet-Uncensored-16.5B-GGUF into shape via the system prompt box. It starts off amazing, customizing the skeletons of the character prompts I give it, fleshing out the basic environment in logical ways, and, if fed detailed multi-sentence prompts can keep the narrative flowing right up to the point where my characters start interacting physically. Then it has a tendency to get in a deep rut of samey responses repeated endlessly. At this point I’m not certain its a flaw with this particular model, or, more likely, I haven’t found the magic balance of prompts to keep it’s context window open to variety.
Dark Champion is my fallback model, but it has several ‘experts’ in it that quickly start lecturing about explicit content. They can be over-ridden by refreshing the prompt or “consenting” to dubious content via a direct message to the LLM, but it’s still a bit of an immersion breaker, even though it outputs wonderful stuff. Additionally it has a tendency to have characters start falling in love in the most cliched way possible. I mean, in my real life, sure, give me intimacy and trust and consent and a fascination with my adams apple (the training data must be made up of 7gb of soft-core romance porn focused on body hair and adams apples), but, dammit, can’t these people just be horny for a second?!
I’ve tried a mess of other models and they do ok, and the smaller instruct models often adhere to the game output better, but the DavidAU stuff seems to flesh out my worlds the best.
tldr: DavidAU makes good stuff. Dark Champion is creative in fleshing out a scene, seems to take longer before falling into an intractable narrative rut but wants the characters to stare into each other’s eyes from inches away and build ever increasing levels of “intimacy.” Darkest Planet is initially amazingly creative, horny as hell, but seems prone to locking into narrative ruts the minute characters start to rut.