the new AI makes spelling mistakes sometimes??? like sometimes it forgets letters or even entire words.
Every choice transforms your body and shapes your adventure · By
this is a common issue with smaller AI models like mistral nemo, you should use a bigger model such as one of the free models on OpenRouter. A quick fix is to edit the AI message to remove the redundancies (if you leave the repeating stuff in, the AI will look at your message history and likely write more repeating stuff later on…)
FieryLion summed it up pretty well. I've played with models under 20b parameters mostly, and they repeat themselves *frequently*. They'll basically blueprint or template part of a message and just keep pasting it back in like it's boilerplate. I had pretty good success with a 30b parameter Gemma fork running locally (I'm on a 2070 Super with 8GB VRAM), although I only get 3-5 tokens per second (super, super slow). But the dialog quality was absolutely stunning compared to the 8b and 14b models. Qwen 2.5 is listed in several places as a "potato-friendly" model, but it's very unsophisticated: it can't really handle lewd/suggestive content in anything that would titilate or suspend disbelief, it is very repetitive and subject to "GPT-isms", and even if you set it for high temperature and sampling (to add a lot of randomness and creativity), it's still heavily constrained by how small it is.
I _personally_ don't like running AI through a hosted endpoint, even though I could be using some really good hardware and getting really fancy models out of it. My two main reasons are privacy and cost. But if you want better quality responses, you're going to have to either pay for much better hardware or hosting solutions (such as OpenRouter).
I've found that it's not always reporting an error properly in the LM Studio when the AI's context length just barely isn't enough, and it goes off trying to generate a response anyway. This is what's caused spontaneous scenario resets for me, mainly on very big/detailed worlds.
Try increasing the context length a little bit and see if that changes it.
Yeah, basically what FieryLion said. Even the web version's AI is dependent on a Context Length set on the ai host, though in the web version you'll obviously not have access to it. (For reference, Context Length is simply how much data the AI is configured to be able to handle at a time in a single request, action prompt + system prompts + world rules/stats/entities/location data etc)
The fact it happens on all worlds for you seems unusual though, so my hypothesis is uncertain there as I've only experienced it on heavy worlds.
Basically, my hypothesis was that if say, the AI was configured to 4000 token context length, and a world required 3950, the buildup of memory containing history, or the Notes, could push it past 4000 in an edge case the AI isn't configured to react to since, just maybe, it only checks token limit BEFORE taking into account history, causing it to discard all past history and reset the scenario to fit within the configured 4000 token limit. But eh, I don't really know the inner workings enough to really say anything certain, which is why I'm just hypothesizing something that could logically explain what's happening, since increasing context length HAS worked for me.
it looks like your context length might be wrong, check your context length setting in LM STudio. Make sure context length is larger than the in game setting for max output tokens + max memory
All right, seems to work. So this max memory, is it the previously generated text? Do the world settings (locations, descriptions, etc) consume it?
Also, would you answer a few more questions i have about how the program works? The definitions for game world, stats, entities, locations and dictionary, are they sent with every prompt or only at the start of the game?
did something happen