Post by es4rtg in v0.0.1 - Known Bugs, feedback, suggestions

Viewing post in v0.0.1 - Known Bugs, feedback, suggestions

What open source model did you use, and were you able to get it to make a new game? I have 24GBs of RAM, but I'm unsure what backend to use and with what parameters as I keep getting a connection lost mid way through or something about the generated class catalog was not usable. Failed: the class catalog

Would be really cool if in future you could add additional context when creating a game to specify certain elements, or modify the context of existing characters

Additionally to force end a scene as I'm stuck in an infinite loop essentially, retry a scene if the ai messes up too

Venus Dev2 days ago

Hey es4rtg, I don't use local backends to play, I only use Gemini Flash 3.8 through Google AI Studio. If you're getting generated class catalog was not usable, more likely than not your local backend is duplicating class codes for certain classes. This can be an intense task for most models since it generates dozens of classes that all need their own unique class code. I've recommend to other users to use a frontier model for starting your game, then switching back to your local backend once you've scheduled your classes.

Character profiles are coming in the next update! I've already played as a celebrity and a talking cat.

Usually I'm able to "force end" a scene by saying that I'm going home or something. The LLM is supposed to classify the scene as over and end the scene. If this isn't working, your model is likely not getting the hint that it should end the scene

es4rtg2 days ago (1 edit)

That's really good and looking forward to the next update, it's a brilliant project so far!

A few other things - Is 127.0.0.1 hard coded? I can't seem to get it to work over my local network but it works fine on the same device (Got it working by opening a local proxy on my client device - netsh interface portproxy add v4tov4 listenaddress=127.0.0.1 listenport=5001 connectaddress=(HostIP) connectport=5001 - & then using http://127.0.0.1:5001 to connect (KoboldCPP) which worked

Hopefully you've changed your API key as there seems to be one for Gemini in the .json file, unless that's a placeholder

Definitely worth adding a re-swipe since I can see what you mean by local are lackluster compared to Gemini, I found Gemma 12B IT Heretic GGUF Q6 to be quite good and it gives me the generation speed, but it does mess up once in a while so that & a force end would be amazing

One other bit of feedback is allowing the option to keep memory of previous events or changing how pre-determined events are created, maybe having it done when the scene is created rather than being statically baked in the initial generation - For example, if you've been talking about something specific with characters the morning/night before, and then go to visit their dorm, one of mine was pre-written text about them never meeting me before (Gemini Flash 3.6 Lite)

Again really good project, keep it up

Venus Dev2 days ago (1 edit)

Um, please send me an email or private message about the Gemini key in the json file, I think I know what you're talking about but if I really did expose a key that's definitely NOT intended and I need to remove it asap.

I didn't hardcode anything for the endpoint, but we do add /chat/completions to the end. This seems to work for koboldcpp, I didn't find any backends in my quick research that this would be a problem for, but if it's a problem for your backend definitely let me know.

You can currently re-swipe by loading the last autosave and typing in your action again

In the next update, you'll be able to manually edit character memories! Stay tuned

es4rtg2 days ago(+1)

Hi again, I've sent you an email

The issue seems to happen on KoboldCPP with a header error, but after doing that fix, it works seemlessly on KoboldCPP & Llama.CPP, if someone else has the problem then that fix is there readily available

Hopefully the new update is out soon as it sounds like really big improvements are coming, I'd love to see the ability to modify temperature and other parameters to adjust outputs as I'm now using Qwen 3.8 27B, it works exceptionally but the dialogue is very strange

Venus Dev2 days ago

Glad to hear it's working! Aside from temperature, what other parameters would be useful? I hate UI bloat and opening the koboldcpp generation menu gives me a headache. I'd like to avoid that and only include the useful stuff

es4rtg1 day ago(+1)

Definitely frequency & repetition penalty modifiers would be great to see, maybe a system prompt input to further help direct a local language model to help it write a certain way, or ensure it uses sprites, change backgrounds & actually end the scene

top P & k too, but that might be a stretch as that's going into ui bloat

Also the Gemini Key problem was something on my end so that's of no concern and that's my bad

Venus Dev1 day ago

No worries! And thank you for your feedback, I will add it to the dev plan

itch.io

Viewing post in v0.0.1 - Known Bugs, feedback, suggestions