how do i make this do dnd 2.5
Google?
>>739493956have 128gigs of VRAM at minimum with another 128gigs of DRAM
this shit is so ass, I have no idea how the retards on aicg are so easily entertained by the stuff this spits out.
>>739494637>he hasn't pledged loyalty to the Celestial Kingdomlocal is ass
>>739495302A big part of the quality you get is how you prop up the engine with instructions. In other words, your "main prompt." You can patch pretty much any weakness if it's a strong model. Because the LLM will do its best to follow your instructions.
>>739495806+50 palantir points, keep shilling good goy
>>739496168I'm talking about open weight Chinese models though. Why would you use Israeli-American censored, overpriced and proprietary garbage for roleplay?
>>739496356local will always be king
>>739493956>Learn to use cards on chub.ai>Then learn how to make your own cards>Then learn to use ali:chat + plists for your cards>Then realize lore books and author notes are better than cards.That's the real black pill
>>739496609>chub>Beginning on the 15th of May, Chub will start banning characters, lorebooks, presets, and stages involving individuals that are, or "appear to be," under the age of 18.I hope you already downloaded any bots that interested you. I've already had at least 2 of my characters deleted and they weren't even underage.
>>739495806I personally think I am traumatizing the Deepseek interns that read my chatlogs, lol.
>>739496708I make my own bots, I only use card to make a summary about the world, lore books have all my characters, locations, races, etc. I make my own jailbrrak text. Deepchink is fine but I liked Mistral Small Creative even if its format is shit.It works on my machine.
>>739496708the cards are so low quality they arent worth saving. make your own
hate the UII'm making my own client insteadit will have multiple tasks per turn like plan, brainstorm, rewrite, summarize etc.also I will finetune TTS from VNs
Alright, i'm trying it.What local model do i run?Do they all just accept lorebooks/characters from chub?
>>739496919i'll make the icon
>>739496890On average, sure. But you don't have to settle for "average"... you have the ability to pick through them for the select few genuine good ideas/executions. I still tweak every bot I download though.I still make my own bots from scratch too, but... I always lose interest in actually chatting with them when I spend ages sculpting an idea from nothing.
>>739496708>chub cucking outWhat will people jump ship to this time?
Do not use chub.Lore cucked out to the EU>>739497119botbooru right now
>>739497057Directors don't enjoy watching their own movies.
>>739496919https://huggingface.co/bartowski/google_gemma-4-31B-it-GGUF/resolve/main/google_gemma-4-31B-it-Q3_K_M.ggufand run it with this https://github.com/LostRuins/koboldcpp/releases/download/v1.113.2/koboldcpp.exe
>>739495302It's entirely dependent on what model you're running, so your comment doesn't make much sense. Or on you knocking on the front end specifically?
>>739496919>it will have multiple tasks per turn like plan, brainstorm, rewrite, summarize etc.mite b cool
>>739496989I use qwen 3.6 35B A3with 24GB gpu
>>739493956it can't at least it couldn't last time I tried, it can't even remember stats. most models suck at contextual memory, maybe something changed but i doubt its public
>>739497137>botbooruHuh, neat. Thanks.
>>739497407dont use moes, get the 27b dense. it's smarter. moes at 3-4b are useless. same with grmma, use the 31b
>>739496989All characters are compatible with all models. All a character card is is a list of attributes and a loose description of a scenario. If your LLM can read English, it can figure out how it's supposed to play a character. Smarter models will probably do a better job though.
>>739497634>5x slowerhmmm no
>>739496989>What local model do i run?Make account on openrouter, put in 5 bucks, use deepseek v4 pro, official APIOr if you really want your lolishit to stay on your PC, have a strong GPU and like dumber slower model, look up how to set up koboldcpp and use the latest meme 31B modelIf you can't run a model at that size or above lol don't bother
>>739496168palantir? this is a ccp website gweilo pig
>>739496527>local>kingmore like manor lord of some shitty swamp in East England
>>739497686pilotredsun is the man
opposing local automatically makes you a shill, YOUR local sucks because you're not rich, and you're defending people who are. fuck you.
>Writing up some quick bs to fuck a slut does nothing for me>End up spending 3 hours writing a whole setting and character arc.Fug.
What models do people use nowadays? I've just been sticking with the latest Opus since I can pay for it, but is there anything better now?
for me it's scenario cards that you can alter and edit as neededfor some reason I can't jive with character cards unless they're also included alongside the scenario
>>739498417no opus 4.7 is far and away the best model
>>739493956ST has a feature where you can have separate text/lore entries that are pulled into context only when relevantyou'd have to get all manuals as text (preferably separated into sections) properly tag them etc.it would be tedious and complex but I think it's doable
>>739498136The correct strategy is praising open source, with weights anyone can run. You will never run Claude locally even if you have a future supercomputer, because it's closed source Mossadware. But you could run the full 1 billion parameters of a new Deepseek or GLM model if you have or will have the capacity. Others can run it too and sell you access. Freedom wins.
>>739497865>>739496989>official APIuse modelrun and deepinfra as the model providers, they only store prompt input / output in ram so its deleted after the request is served.
>>739498417How much is Opus to payfag for?
>>739495806What model?
>>739499338>use modelrun and deepinfraThose are good ones and don't fuck the model with a bad system prompt, but Deepseek themselves are giving a good discount for the official API right now>>739499749>How much is Opus to payfag for?If you have to ask, you can't afford it.
>>739498302I fail to see the problem.
>>739498302thisIf I want a quick want I will just gen some pics to fap toI need LLM to make virtual gf
>>739502468a quick wank*
>>739499860If you have more money than sense or a way to scrape keys, Claude Opus. I like 4.6 better, but 4.7 isn't bad either. Generally speaking, Opus is going to be your best option if money is no object. Look into jailbreaks, it doesn't take much effort at all to make any sort of LLM do shit that'd make /d/ blush.Otherwise, DeepSeek V4 Pro is almost as good as Opus, and is like 10 times cheaper.
>>739502616Opus used to be really good but Anthropic are busy at work ruining it. 4.7 is notably worse at roleplaying prose and flavor than 4.6, which is worse than 4.5 etc.>Otherwise, DeepSeek V4 Pro is almost as good as Opus, and is like 10 times cheaper.it's 27 times cheaper lmao
I used to shill claude but qwen local is pretty much as good and I can make walls of text brainstorming ideas for free
Lately, I've switched to using FrontPorch instead of SillyTavern. Much more user friendly but not as customizable.
>>739503001Sillytavern is built for ricing, you're supposed to install all manner of custom extensions and vector storage and custom CSS and scripts and regex (some of the add ons you install will steal your keys too, as a treat)If you want plug and play and won't modify anything there are more user friendly options
>>739503084what are some good custom extensions? I've used ST a lot but never bothered with them
>>739503084>extensionsEnjoy having your API key stolen lmao
>>739498302>have an idea for a character with some random gimmick>write out a bunch of autistic realistically plausible explanations for all the different elements of the gimmick>expound on all the ways it might affect the character and the world>absolutely none of this ever comes up in the actual story
>>739503262>reading based hobby>anon can't read half a sentence longer before postingGrim!
Jerking off too much because of gemma. It's like Gemini 2.5 pro when it was free, except it runs ultra fast and it's uncensored and local. It's great for rp but with the mmproj I'll also just spend hours molesting it, but showing characters porn is cool too
>>739498302>>739503447>start writing a quick fairy character because I came up with a dumb gimmickiron drives away fairies, so forcibly make her wear iron piercings and sex toys to keep her weak and enslaved>get so carried away with lorebooks I end up writing an entire fantasy setting with areas, polities, races and magic system>now rewriting it to cut down on tokens
>>739503663I didn't even know those existed, pretty neat, how do you get the characters in SillyTavern to see images? I can get stuff working through Kobold but how can you pass it through to ST?
>>739505036nvm I found itcool stuff
>>739505036I've noticed it do some strange things, like ST will process the image first and identify it in a sfw way (These people are hugging very close. There's a lot of white slime. Etc) but when it reaches the point the character's acting it can tell what's happening. Pretty fun, had one character see a pic with a huge load in it and she obsessed. Any other porn and it was like she was just humoring me until the next big cum pic and then she'd lose her mind
>>739495806>Celestial kingdomWhuh?You mean cloud?Are you using sillytavern? I use openrouter, I better get looking for some jailbreaks too
>>739506125Do you have a specific gguf and mmproj or prompt you use for the caption gen? I'm trying it now and the thing is just generating a safety guideline message and refusing to describe the stuff I put in.
>>739508924That particular time I was running mmproj-Gemma4-26B-A4B-Uncensored-HauhauCS-Balanced-f16 with the gguf quant for the same model, but I've also tried it using a different uncensored quant model (gemma-4-26B-A4B-it-heretic.Q5_K_S ) and the regular mmproj from Google and it seemed fine. Thus far I haven't decided on if I prefer the Hauhau model or the heretic model but for the mmproj itself I'm not sure it even mattersHow I handled it through ST is I would click the wand icon near the text entry area, add an image, add the image, then I'd type in a prompt talking about how I just changed the image I was showing to the character and I'd ask the character talk about the picture and make up her perceived background story for the pic. Never actually had it outright refuse me
>>739508924NTA but for images I use Chat Completion, then in the AI response configuration menu [leftmost option at the top of the ST window] enable "Send inline images."From there you can click on the ... next to your prompt or the wand at the bottom and attach a file.>Analyze the subject in the provided picture and..."
>>739509459cheers, I was using a different gemma 4 and it seems like that was the issue, google's brainwashing is strong>>739509706thanks too anon, I'm just using the wand to make a caption and then trying to have characters respond
>>739495302Chatbots are be very amusing if you are creative.
>>739493956You don't. AI's still in a shit state and prone to hallucinating and forgetting shit so every campaign you try to roleplay with it is inevitably going to turn into an incomprehensible mess with shit like you fighting an owlbear only it's suddenly a goblin only it's suddenly a mimic, only it's suddenly a treasure chest that needs to be lockpicked to open and for some reason lockpicking it is purely a strength check.
>>739503663Gemma is nowhere near 2.5 pro.2.5 was insanely good, and we had it for free for nearly half a year?Now we only got 3.0 flash, which is pretty mid. God, I wish I had a pro model for free again...
>>739510861For RP it hasn't felt too far off for me, and since it can run local it's even better. It'd be nice if they turned the taps back on for free gemini since it's nice being able to just start ST and immediately dive in instead of taking time to stoke the boiler of the infernal cockmilking machine first
>>739510861thisgemini pro could literally follow tabletop rulesets with easeflash can barely remember character details in cardsgemma is even weaker
