It doesn’t matter what you use, OOC is consistent in Ai roleplay, in Ai RPG, Ai chat, sillytavern… Etc and it always worked just fine, but it seems whoever made the story gen had a different idea in how it operates, been fighting the Gen using OOC with 50/50 success rate, it worked and sometimes it doesn’t (very hilariously rebellious), I just keep pasting the OOC after the last written word until it finally caves in, try it.
OOC is one of the tools that everyone in Ai rp communities knows and use, are you sure that you used it correctly? Because I did use it multiple times to fix this exact problem.
Yep, hyperfixation is a one of the things that most LLMs are guilty of, you can easily get rid of it by using one of these two methods (or both):
edit the message, remove the mention of the homework/historic event until the Ai let go.
use OOC in your response, OOC is a very powerful tool that you allow you to do whatever you want in the story, here’s an example:
“alright, we are done with this homework, finally we can switch to other things, yeah?”
(OOC: do not mention the homework that revolves around [historic event] ever again in the story).
It REALLY depends on two things :
1- your model of choice.
2- your specifications.
I have 32gb ram (ddr5 @ 6000mhz), rx 6950 xt with 16 gb of vram and an i5-13600k, the model that I got to run comfortably range between 24b and 27b, anything higher than that got my pc screaming for mercy plus slowdowns and incoherent output. With 6 gb of vram, I highly advise you to run Sao10Ks L3 8B Stheno v3.2, either Q3 quants or Q4 quants, but I highly recommend that you download the Q3 quants (for higher token count), assuming that you have a 16gb of ram, you can run between 8k and 12k tokens, which is higher than the token count that the perchance model have (6k tokens), also, use koboldcpp and connect it to sillytavern, it’s a bit complicated but once you get it running, it’s a very smooth sail.
On a side note, if you could get your hands on a clean card with a lot vram like an rtx 3090 with 24gb of vram on the used market (got one for less than 400 bucks), you can run much, much better models locally, models that would nuke perchance’s Ai chat into oblivion with much higher context windows, like Gemma 4 31b, WeirdCompound family, Cydonia v4.1 24b… Etc.
Aye, glad to help