Talk about whatever u want
Every choice transforms your body and shapes your adventure · By
To wszystko! Teraz powinieneś móc grać w grę używając szybszych modeli AI!
Ważne:
- Darmowe modele mają limity żądań!
- Różne modele mają więcej/mniej limitów. Możesz eksperymentować, zmieniając darmowe modele.
Najlepszy darmowy model: meta-llama/llama-3.3-70b-instruct:free
- Ten model jest bardzo inteligentny i ma ekstremalną pamięć (128K!)
- JEDNAK: ma bardziej ograniczone limity żądań, co oznacza, że nie możesz wysyłać wielu zapytań, może raz na minutę
tbh? If you’re willing to pay openrouter a few bucks you can use the better models there, I’ve tested some
Nah, I will pass on the paying. But was cool til it worked, and now I just use your setup. But can you like, always play it by yourself? I just don't know how this whole system works, to be very fair. I just did as it was in instructions, it then worked some time, because I did it right, I presume, but now the fate hit and I'll be using your setup. Thumbs up still for trying to make it work, though! :)
No worries, I have downloaded the game. That part won't stop me. :V
Also, thanks for the hint about KoboldCPP! I was gonna try it using Ollama (with instructions from our old friend ChatGPT), but since Kobold can also host it as an API server, I would try that method first. That will be tomorrow though, for now I need to catch some sleep.
I was able to get it to work! However, the output seems to fail at parsing \n and it just treats it as literal text. I'm not sure how you run your AI, but do you know a way to change how this works in KoboldCPP so that it formats it differently? It's not the end of the world, but I figure it's worth asking. >.>
did you use the correct url? it should be http://localhost:5001/v1/chat/completions
replace 5001 with your port number
You need to keep these stats for now because they are mandatory for the 3D character viewer: Stomach, Fatness, and Breast size. Otherwise go ham, make whatever you want and share your worlds so others can play too c:
Keep the description of the entities and locations concise because too long and you use up the AI’s limited context limit, or you can use better AI models with more memory
Genuinely, at least what I do, is I type out a description, role, goal, and general capabilities of a creature, take it to the AI I plan on using, and tell it to try an minimize the amount of context tokens used while keeping the theme, goals and capabilities of the creature intact for a Storytelling AI to use, it tends to work fairly well, you do get an AI occasionally going off the rails with it though.
You should use an instruct model because regular RP models fail at following precise instructions like updating game stats, as they are meant to produce flowing story and not manage an actual game with RPG mechanics.
Also I heard complaints from other players that koboldcpp doesn’t work well as an AI endpoint, so you may want to switch to LMStudio instead
That's.... very much a limit of the AI. I might be speaking for the creator a bit, but this is just what I have noticed after several hours of tinkering with the worlds, the AIs, prompts, and how AI context works. Genuinely, there's a few things that can happen, 1 is just that the AI ignores the descriptions of the stats, 2 might be that AI using "Rolling Window" or "Truncated Middle" for context Overflow (IE more data present than the AI can handle) and the AI just starts throwing data out, and inadvertantly throws out the fact that you had the trait, especially if you use Rolling Context, where unless you are specifically telling the AI you have X trait every other passage 9Most rolling windows only have a 2k Context limit) the AI will completely forget you have X trait. Or it could just be that the AI doesn't see the trait as important or have an effect on the story, and again, it just throws it out or ignores it.
There are limits to the AI, but in this case traits are never passed to the model. Using devtools to check the network request sent to chat completions, it shows all of the information that gets sent to the AI. As of 1.0.3, it sends the system prompt and however much of your story that fits in the context limit you set. For me this was set to 64k (model supports up to 128k), but setting it to 1k reduces the information sent to just the system prompt and latest action, so the system prompt isn't being thrown out. I'm sure it wouldn't be handled as consistently if the story context grew to outweigh the system prompt, but that's a separate issue inherent to how messages are weighted. The issue here is just that traits aren't included in the system prompt, so the AI never knows they exist to begin with. Adding "<TRAITS DESCRIPTION>" to the system prompt, similar to how world data and stat descriptions are handled doesn't add any trait information, so there's presumably no code to support it.
Sorry the AI can’t update the min/max or regen of a stat just yet, will be very soon. You can still tell it to increase the pregnancy stat when the conception stat is positive, just change the stat update system prompt in the settings (latest version). The AI will focus more on updating the stats that you describe there.
You kinda have to realize the majority of "good" AI's nowadays are those that recieve constant updates, which tends to mean large corporate AI's that have a lot of restrictions because said corporations really don't want their AI's doing anything that might get them sued, and even more so if kids can easily get around it the blocks put in place. So you just kinda have to accept, large AI models like llama are going to constantly have blocks being put up, and it takes time for people to get around those blocks. So genuinely, don't be surprised if your experience is very intermittent due to said updates, it's just kinda part of the ecosystem.
I've had the same problem after the update, I tried to make instructions for AI to ignore it's filters of *Unethical content*. But after an hour and hafl I gave up xD
I just fixed it by using diffrent model: Mistral-Nemo-Instruct-2407-GGUF to be exact. With it problem is gone and everything works fine although it is a lot heavier on my pc than Qwen2.5 that I used prior.
So Not wanting to speak for the Dev, which i know already replied, but it does help to run your own AI locally if you have the computer to do it, there is a thread on how to do it pinned in the community here. It saves a lot of headache if you can, and if you can I can get about 13k context tokens (IE roughly 52,000 Characters, 1 token = roughly 4 letters) on 32 Gb or RAM and a RTX 3070 laptop. I am aware it's a bit higher spec than what most people run, but the 13K is all run on the 3070, which is about 12 GB in RAM. Would I suggest flooding your RAM for it, no, but the game, at least when going through the story, seems to only log and send roughly 2K context memory, so 8K characters, and the rest is used for things like System Prompt Additions, Stat entries, Entity entries, locations, and traits. so that is, the story itself, and agin, this maybe wrong, it's just what I have noticed when I created my own scenerios, only keeps about 2K in context memory, and if it does, Dev man, you are amazing, I love you, all I want for christmas is a way to adjust how many game context tokens you can tell the game to send to the AI. Within reason, I know you're already down a kidney.
all I want for christmas is a way to adjust how many game context tokens you can tell the game to send to the AI
Its in settings, memory limit (not that the memory limit here only counts the game events and the world data is on top of it)
In the settings of the game, the token output what exactly does that do?
It determines how many output tokens max, more max tokens mean the AI write more and more text
ahhh I see. Something else I noticed and I just figured it’s the way AI work but when I type in a command. Like say “look for a water source” I notice that the AI will say I find a water source and then drink from it. If I don’t want the AI to make me do something do I need to be more specific with how I explain what I want to do? Also sometimes when it generates the text it will only be around one sentence long. Does that have something to do with my internet connection? The settings I have in game are the default where it says paragraph, essay style. I have not touched that stuff.
Yeah, sadly the AI will assume that, if you look for food or water, you will automatically want to eat or drink it. What I have found to help is to specify in the game text prompt that the AI cannot take actions for the player. If the player wants to find water, then the player will ONLY find water and do nothing else until specified further. Without fixing it too hard (so that the AI does allow some new events, like entities, to stroll in), you could add something like this to your game prompt:
The response can't assume or take any actions in behalf of the player.
for now you have to write in the world rules to tell the AI to update the stomach stat when pregnant. I’ll add the ability to link stats so a pregnant stat can be linked to stomach stat and such
