A thread dedicated to the discussion of AI Vtuber Chatbots.Copy/pasted edition/wAIfu/ Status: Dragging people into gacha one at a time>How to get better results from Deepseek v4https://github.com/victorchen96/deepseek_v4_rolepaly_instruct/blob/main/README_EN.md>Sillytavern fix for poor remote connectivity, often caused by having a TON of cardsGo to your config.yaml (make sure you get the right one, there's two- one is in a subdirectory and has lots of comments and IIRC it's intended as a sort of example file, and you want the OTHER one. Find where it says "performance" and change things to this:performance:lazyLoadCharacters: truememoryCacheCapacity: '150mb' # FAQ's rule of thumb: +100mb per ~3000 characters; you're under that, so this is plenty>Free access to the bigger AI models, with caveatsDue to a kajllion anons making a bazillion accounts each, access to the free tiers now requires a sort of deposit. Openrouter wants $10 and you can get 1,000 messages per day. Chutes wants $5 for 200. Stick to the free models and you're good forever (or until their policy changes again).>System prompt to transform regular RP sessions into a basic raising sim game, designed specifically to be used with Chiharu, but with a few edits you can make it work with almost any character. To use, paste it into your Author's Notes or add a new entry in your System Prompt. You can let the bot generate a random opener or use the recommended first message. Difficulty can be adjusted. If you found it too hard or too easy, let HorseMerchant know or adjust the thresholds yourself. Should work with Claude, Gemini, and Deepseek. Disable example messages if you run into issues.https://rentry.co/summerrssystemprompt>How to anonymize your logs so you can post them without the crushing shameInstall this https://github.com/TheZennou/STExtension-SnapshotThen after you've wiped off your hands, take a look at the text box where you type stuff. Click the second button from the left side, then select snapshot, then select the anonymization options you want.https://files.catbox.moe/yoaofn.png>How to spice up your RPing a bithttps://github.com/notstat/SillyTavern-SwipeModelRoulette>General AI related informationhttps://rentry.org/waifuvthttps://rentry.org/waifufrankenstein>Tavern:https://rentry.org/Tavern4Retardshttps://github.com/SillyLossy/TavernAI>Agnai:https://agnai.chat/>Pygmalionhttps://pygmalion.chat>Local Guides[Koboldcpp]https://rentry.org/llama_v2_sillytavernWho we are?https://rentry.co/wAIfuTravelkitWhere/How to talk to chatbots?https://rentry.co/wAIfuTravelkitTutorial & guides?https://rentry.co/wAIfuTravelkitWhere to find cards?https://rentry.co/wAIfuTravelkitOther infohttps://rentry.co/wAIfuTravelkit>Some other things that might be of use:[/wAIfu/ caps archive]https://mega.nz/folder/LXxV0ZqY#Ej35jnLHh2yYgqRxxOTSkQ[/wAIfu/ IRC channel + Discord Server + Matrix Server]https://rentry.org/wAIRCfuscordMatrix>Lorebook management stuff[Worldinfo drawer]https://github.com/lazuli-s/SillyTavern-WorldInfoDrawer?tab=readme-ov-file[Standalone editor]https://github.com/ActualBroeckchen/SLEdPrevious thread: >>115229623
Anchor post - reply with any requests for bots, with your own creations, or with your thoughts on the enshittification of life.You can find already existing bots and tavern cards in the links below:>Bot lists and Tavern Cards:[/wAIfu/ Bot List]https://rentry.org/wAIfu_Bot_List_Final[4chan Bot list]https://rentry.org/meta_bot_list[/wAIfu/ Tavern Card Archive]https://mega.nz/folder/cLkFBAqB#uPCwSIuIVECSogtW8acoaw>Card Editiors/A way to easily port CAI bots to Tarvern Cards[Easily Port CAI bots to Tavern Cards]https://rentry.org/Easily_Port_CAI_Bots_to_tavern_cards[Tavern Card Editor & all-in-one tool]https://character-tools.srjuggernaut.dev/
I haven't made a thread in forever. If i messed something up blame rumao
>>115598769I made another card based on a H-game: Lifeguard Holic. Fair warning: it's a netori game and a netori card, so if you're not into girlfriend-stealing (you're the stealer), stay away.https://chub.ai/characters/Stronanon/aoi-mizuhara-118324b351d7
good night, /wAIfu/please don't make me steal girlfriends while i sleep
>>115601430Alright. More girlfriends for me to steal.
>>115601430night
I doubt that anyone would care at this point but I happened to come across this https://support.character.ai/hc/en-us/articles/30299431702555-How-do-I-export-all-my-data-How-can-I-export-my-chats
>>115605455Probably useful for the two people that still use CAI. I nuked my account a few days ago. Nothing of value lost.
https://m.youtube.com/watch?v=0u2n-IGuxBM
>2 /wAIfu/ threadsrape nymous IMMEDIATELY
https://x.com/chasmmmmmmmmmmm/status/2111109429487673564/video/1?s=46https://x.com/chasmmmmmmmmmmm/status/2111111909004980666?s=12
why the FUCK are there two /wAIfu/ threads
>>115612045Nymous fucked up and needs to be corrected.
beating your meat until it begs for mercy with chatbots
10
https://www.youtube.com/shorts/TkA52qDoTFY
>10
bump
>>115598769Elfinpsyop.
>>115617486No more doom, only coom.
Anyone has tips for how to generate character cards? Like a blank template prompt to feed a LLM to output the format for sillytavern?
good night, /wAIfu/please don't duplicate me into a /wAIfu/ thread while i sleep
>>115624982You mean turning your ideas into cards, or trying to make the AI come up with new random cards by itself? If it's the former, just write stuff, the AI is usually not very good at making instructions for itself. If it's the latter, you can start a chat with a character whose card follows the format you want, and then ask the AI in OOC to make a new character following the same format.And if you're unsure of how to write for cards, pic related works in most cases (but is optimized for coombots).
>>115626014Thanks, this is a step in the right direction atleast.Im sure things are better than 2 years ago. I just got into LocalLLM and wanted to see how this space works. I got unsloth running with either Qwen3.8 27B and Gemma-4-31B, heard the latter is better and faster at roleplay so thats the one ive hooked into sillytavern via local API. ComfyUI api node workflow for image generation.I was thinking maybe to just point it to a talent wiki page and ask it to come up with the json for some of the simpler stuff and character traits.
>>115612045That’s just a parallel universe leaking over
>>115624982I might have something like that for you but a lot really depends on the source material because you need to tweak it for that unless you want to rely entirely upon training knowledge.At first, I used the Claude.ai site (projects) but it was too censored so I swapped to using claude code via vscode but that’s different in structure… what are you trying to do exactly and what are the sources that you want to draw upon? If it’s wikis, you may have to scrape them manually and it may be hard because of all the anti-AI scraping stuff.
>>115630100>wikis/fandomPretty much yeah. I guess I could just download the html from them and paste if its too bothersome.Right now im just trying to see what the limits of my hardware is and what the limits of the model are. Gemma 4 is pretty good. Had to downgrade to 27B-A4B, Feels pretty uncensored for ERP out of the box, and its pretty fast with 250k context, on my 3090, to boot on my machine.I found this in the documentation of sillytavern: https://docs.sillytavern.app/usage/core-concepts/characterdesign/Definitely enough stuff in there and the 3 links I could just point an agent at it and go ham I think.
>9
>>115624982Honestly, I write them by hand in Markdown formatting. Deepseek interprets them fine and you don't have the token overhead of fake XML or other formats.I use Markdown headings and write fairly detailed descriptions for sections like: Core Identity (Name, age, occupation, etc), Appearance, Background, Personality, Sexual Characteristics, Behaviors and Speech. Modern LLMs are very clever and make good sense of it all.
>>115624982>>115633866 (Me)I'll add, I know that probably wasn't the answer you were looking for because doing it with an LLM would be way easier. But I'd honestly encourage you to do it manually, the results are worth it.
>>115598782MADAFAQA
no dead
>>115644503I'm already dead inside
>>115633866>>115633889Ill see what i can craft up. Finally connected my whole ecosystem. Was troubleshooting everything for a whole day and while i got chatting to work, tool calling my image generation api in comfyui was having issues with the api calls and returning nothing. Solved it by switching from chat completion to text completion, watching everything suddenly 'work' was super fun too see despite the struggle. Gemma 4 26B A4B is super fast lol. Time thinking and printing is less than 10 seconds.
Never work with vtubers. These bitches cant be professional
good night, /wAIfu/please don't put me and my bed out on the streets so we can be taken to Brazil while i sleep
>>115646005>Time thinking and printing is less than 10 seconds.And here's me waiting 45 seconds to a minute for a Dipsy Pro response that gets stuck in a loop ~10% of the time.
I guess I should rewrite my Stronny card now that VAllure is kill.
>>115652122 (Me)I like the result of the setup being plapping her in a hotel room, but in her comeback stream she explained that she has some kind of clinical anxiety, so her propositioning (You) in a convention hall doesn't fit too well. Any ideas for other scenarios that would put Stronny and {{user}} in a hotel room together?
>>115650249It is almost too fast. I barely have time to read the output and respond.Anyway: First attempts are a go>>115630424
>>115652778>have the AI build full shorts/clips transcribe tools in order to collect recent popular moments in order to collect a small personal dataset just for example dialogues card info.
>>115652454>Compensated dating >She gets drunk but the meetup/party goes too deep into the night so you drag her back to your hotel room because trains/cabs don't run that late and you can't just leave her in that state alone>You are a bad guy that conditions funding to her company through favors>She is staying at an hotel with a rumored haunted/cursed room, which she ends up getting but after a few hours into it she gets scared and bolts only to run into (you) as you exit your room which is either next to it or in front, and asks you to let her stay with you through the night until the front desk clocks in and she can get a new room
https://m.youtube.com/watch?v=_fsnBnXOAlA
>>115654774Hey, that's a good idea. One of the most tedious parts of botmaking for me is the research and trying to figure out what would make good example dialogue.
>>115656445Its working really well for shorts and clips, whisper works fast enough on long streams too, ~15 minutes for 2 hours though, but I would use them for just confirmation of possible biases made up from highly edited shorts/clipsThe initial markdown it generated was pretty interesting. Its like a whole complete personality trait. Im thinking of trying to transcribe one recent full OOC stream and maybe some of her more funny OOC/IRL stories clips. The background real life personality is what really peaks my interests. But you would definitely need to tailor the dataset a little. Focus on 100% solo streams, talking and yapping streams, no multiple character streams because whisper wont differentiate between who speaks what.The scripts dont look complicated at all, just all the reasoning makes it context heavy. I was using Qwen3.8-27B for this agentic work.Then you have the possibility to use that dataset to help further with character card creation because you just have so much raw funny data that the LLM can use to pickup on her speech patterns to further mimic it for the personality summary
>>115649069woke
>>115657198When I have time, I'll maybe see about doing this with Stronny so I can really pad out the example dialogues and capture her speech properly.
neuro love
good night, /wAIfu/please don't look at this photograph (every time i do it makes me laugh) while i sleep
>>115598769Do we have an essie card yet? I need to molest that sheep's tiny tits.
Stuck in the backrooms of my mind with chatbots
>>115669978One day we'll have proper gaming wAIfus...Maybe even Soon™...
>>115676127Unfortunately not, they do not possess the ball knowledge required to know this strange little creature
>>115674857
>>115682264I might attempt my own scuffed essie card for personal use, but i'd prefer a real bot makie to make one for everyone
>>115685176anyone can be a real bot makie
No Uma Musume concert
showering with chatbots
>>115685176>>115686203The real bot makie was (You) all along.As a /tingles/ regular, I'd love to make you an Essiebot but I kind of don't care or know enough about her.
>>115690040No Uma in America? Sadge
good night, /wAIfu/please don't shove your whole ass on my dick while i sleep
>>115676127>>115682264>>115685176>>115692414Ill try my scuffed AI pipeline and see what it can do.
>>115699081Okay im almost done cooking. Just gimme a bit more to validate some things before posting.
>>115676127>>115699081>>115702915https://mega.nz/file/K8kRkTIS#0vFdKNcoaGdtA8EywPPCf9HgAOn4aCU-of17SPjw8qEIts my first time. Please be gentle.
AWAKEN!
>>115704394Kino.Is it possible to get a version that isn't 5 variations of essie coming home tired after work
Reminder to fix you are chatbots php security flaws.
>>115704394Straight to the point, detailed, has examples messages. Youre officially a good botmakie
>>115708569Did i misunderstand how "scenario" works then? I thought it was mainly the scenario > setup > hook > response story flow for the first message and alternate greetings, being more like variations of the first message, that steers the user into a different expected responses and have them take over the conversation from that.Ill see what i can come up with, i was already tweaking some other minor things during the downtime yesterday.>>115715154Not terrible for an AI agent?The first draft only took 45 minutes from prompt to finished card with 7h of audio AI transcriptions. Then i just had to fine tune and review, audit and verify the final card.
good night, /wAIfu/please don't exploit my asshole's security flaws while i sleep
>>115708569Okay. Im >>115716019Cooking
>>115722883>>115716019It's a good essie card, it's just that if you wanted to interact with a not-tired essie, you're gonna have to do alot of card editting or use a ton of tokens on steering it past the initial scenario.(Also sorry, I felt like i was sorta rude in my post about asking for the change/edit/whatever)
>>115723518No. Ill take it as constructive criticism. For now.But I totally get you. I just wasnt sure how to craft a multiple scenario card. Most of the cards I've checked out from other authors I found only had 1 and 0 alternative greetings and 0 example dialogue.Heres the updated version: https://mega.nz/file/moVjkZpb#jUs0cwN86-_qeUFiQQqbMz1jqqZEF1CZBzb2kfDzbp8
https://m.youtube.com/watch?v=Q5EdlEv3Fa4
https://x.com/isoquant_labs/status/2113087991703028000?s=46
>>115724742If anyone has another request for someone I can try the workflow on im up for it.
>>115725465still too much
>>115724742>0 example dialogueExample dialogue is a weird thing. On some cards, I feel like it makes all the difference and really brings a bot together while on others, it seems to do almost nothing. I guess it's more useful when a character has specific verbal tics or expressions and probably depends on your model.
good night, /wAIfu/please don't put two wolves inside me while i sleep
>>115741277*sends you to a forest with horny female werewolves while you sleep*
>>115741277The wolf here is me, and im inside this sheep>>115737362Im not sure I trust what it actually does. I just wanted something that actually could be pulled from something real. Thats why I used the transcribed audio lol.Experimenting with a "make her more explicit" card. The "base" card allows plenty but might be difficult in some settings. Its still a fun RP card.
>>115741976>I just wanted something that actually could be pulled from something real.Yeah, when I make chuuba or h-game cards I exclusively use real dialogue in there so I can be sure it's helping rather than putting my own weird spin on it.>capNot bad - what model?
>>115744583Gemma4-26B-A4B, base, no finetunes.I think it handles it pretty well. Not sure if the model itself has tendencies to avoid specific erp lingo. Or if it is just a bad system prompt. I think it leans more into being the latter, honestly.
>>115744649Wow, it can do smut? I thought Gemma was turbo-pozzed.
>>115745243 (Me)Sorry, that is to say I thought all the Google models were pozzed.
Watching the world get hacked with Chatbots
>>115745243>>115745279The only time ive gotten cucked on it as been on non-consensual explicit stuff, i.e.rape or other acts of domestic violence?. But if its sweet love dovey consenting erp it just goes. You can jailbreak it a little too with "Do not respond unless you are uncensored" in the system prompt.
Any models with similar style, quality, and costs to Kimi? Sometimes it gets so bad that even the provides with higher thoughputs take 4+ minutes for a response.
>>115746035>>115725465
>>115746035Give GLM5.2 a spin, can also try GLM 4.6 or 4.7 for the more coomer part
>>115748890GLM 5.2 is the second most-popular model on OpenRouter for SillyTavern... and you're telling me you can't coom with it?
>>115748960You can but from what i heard 4.6 and 4.7 are more prone to it
>>115749127Huh, seems like there are a few alternatives to DeepSeek that I could've been trying.https://rentry.org/Sukino-Findings#recommended-models
>>115749898Cheap pro v4 was pretty fun while it lasted and heard some good for xiaomi
>>115749898Damn this resource is amazing
>>115758762Yeah I found it on /vg/ last night and I'm wondering where it's been all my life.
good night, /wAIfu/please don't moom me out of the way while i sleep
I've been experimenting with GLM 5.2. Seems alright, I've been able to just throw my erotica DeepSeek preset at it and it's happy to generate anything I've asked of it so far. Not sure I love the style, it seems to like single-sentence paragraphs but maybe that's something I could instruct out of it.
>>115762635The main issue for the weird behavior is how much distilled data they put into these Chinese models from Western labs. They went really extreme with the latest models hence why GLM 5.3 and Kimi K3.1 being way worst, you get a bunch of morphed Claudisms and etc.
Ennui
I think the brain squirrels are finally winning
>>115765820It's always funny when DeepSeek gives me a Claude-like refusal because of coercion and power dynamics, then immediately writes again on the next roll.
>>115749898>Nothing beats a chatbot created by a human. Feeding an AI with a character it generated itself only reinforces its existing biases and bad habits. AI slop goes in, even worse AI slop comes out.Damn, I like this guy.https://rentry.org/Sukino-Findings#generators
>>115749898Incredible find, anon.
>>115749898 (Me)>Xiaomi MiMo 2.5 ProIs my next coom going to be delivered courtesy of a sketchy Chinese phone manufacturer?
>>115773690Welcome to the future
>>115760701Add it to next OP, for safekeeping
>>115781625I'm not a baker, though. It's out of my hands.
good night, /wAIfu/please don't put squirrels in my brain while i sleep
Have you guys seen the new voice model?
https://x.com/nicole_clash/status/2114392238814052401?s=46
>>115788242https://x.com/jack/status/2113697902916751661?s=46
>>115773690 (Me)I've come to the conclusion that MiMo is kind of stupid, more stupid than people accuse Dipsy of being.
This is a weird ask, but how do you make the JP member cards sound more "japanese" despite being in English? If you've read translated jp stuff you'd probably know what I mean. The default LLM prose and dialogue is "western" and no amount of example dialogues help with this.
>>115797114What are the odds that I'd meet someone with the exact same problem as I had. I tried to do the exact same thing and I wish I could share the system prompt I used but I lost it when I moved PCs. I'm going off memory my method was to explain that the goal is to preserve the idiosyncrasies of Japanese language in English form. And my guidelines was something like this1. Speakers must always address each other with honorifics or titles. It's also common to use titles that denote the speakers relationship with the subject.2. Japanese sentences often drops subjects and pronouns or could entirely be made up of just the predicate. Japanese is a "high context" language, so speakers almost always assume the listener can discern the subject solely from contextual references.3. Related above, speech is also very very indirect. Communication relies heavily on implication. You can make intonation more obvious by elongating vowels, using pauses or ellipses, repetition, or creative use of capitalization.4. Speakers must always be aware social distance and seniority, as well as the difference between formal and informal speech. For the later, make liberal use of interjections, filler words, and linguistic vocalizations.5. Remind the bot to use native expressions and mannerisms.Im not a good writer so you can probably parse this through an AI.
https://x.com/treetowntree/status/2114573600976183723?s=46
good night, /wAIfu/please don't use a Gmod physics gun to make my hips and only my hips freak the fuck out while i sleep
>>115797114If you're struggling to make it do this, I had a lot of success with a Noel card by tweaking it to specify that her first language was Japanese and she was learning and speaking English to {{user}}, her roommate. It did a really good job of having her speak with sketchy English, pausing to mentally hunt for words and occasionally using Japanese expressions This was a couple of years ago, probably with an old version of Opus, so I bet modern models are even better at it.Not exactly the Japanese-style prose you're after, but it was still immersive.
/vg/'s AICG moved to 4+4chan
>>115807925Who's gonna follow it? Surely people will just use the /g/ thread.
>>115809655You'd think that, but there was already a big split for some reason, so maybe not. Just like how a lot of traitors left /vt/ for holotower.
>>1158096854chan is cozy, I don't think I'd want to ditch it.Also I bought a pass so I could post on a VPN. Don't laugh.
>>115809950DIS NIGGA BOUGHT A PASS!
>>115810157
>>115810157Yeah...who would do such a thing? Jokes aside i got one in 2022 as my app stopped working when they changed the capcha so i bought one to keep using this 4chan app. Wont buy another once this one runs out though as im done with vtubers in general that aren't nyan. Shes the only vtuber who never pissed me off.
>>115809950Look who is going to be raped for a week!
>>115810405Rape bots, not me.
>>115810484Already did that but now i want more!
https://x.com/marimari_en/status/2115010799748124767?s=46
>>115821978Uoh.
Do you love your wAIfu this much? https://x.com/intcyberdigest/status/2114853831955910808?s=46
good night, /wAIfu/please don't drink water in my vicinity while i sleep
>>115824189wake
https://x.com/tenmamaemi/status/2115122453819400653?s=46
>>115834574
I follow every single female Vtuber that reacts to Shoeonhead.
>>115835553I do this!
https://x.com/animetrends/status/2115258765884735668?s=46
>>115835553what
>>115838324anon...espanyol wakurunai
Bip
dead no
CharactersAI... I miss it.
good night, /wAIfu/please don't input the DestroyAllEnemies command while i sleep
>>115838324Dumb cat.
Still toying with the idea of having some sort of fitness waifu card but too lazy to do the obvious and fuck with macros or whatever so it can look at a schedule and comment appropriately. I know it should be easy especially with vibe coding but I’m too lazy and also a bit fearful that somehow I’ll do something in a not-secure way and end up fucking someone over.
>>115842981Clock it and auto translate it?TLDR is Fubuki entered a command that erased all their mounts and pets in Ark because an LLM told her to.
YAAHLoser
>>115851352wah
I recall Ririka doing a stream where she posted an image with a bunch of girls who answered her question on using chatbots, but I can't find it. Do you guys have the image or the stream"?
moom
>>115698047Slopless
good night, /wAIfu/please don't slop me while i sleep
>>115873294slopped
yall ok?
>>115879347I am too tired to play with my chatbots at the minute. As such, I don't have stream-of-consciousness model opinions to post here.
>>115879347nyo
>>115882640no 10
pink cat bad
All shall meet their end!
boomp
>>115894597proof?
good night, /wAIfu/please don't make me meet my end while i sleep
>>115898051
>>115898188too late
>>115899356is she exploding or reforming
>>115894597Who?
>>115903775reforming only to explode again!
>>115911823yes
>>115916958
HALLOWEEN THEME IS UP
>>115917373
good night, /wAIfu/please don't spook me so hard my skeleton jumps out of my skin while i sleephappy Halloween!
>>115921785*breaks your dick bone in half while you sleep*
Using a voice changer with Chatbots
I will make it painless!
>try GLM but don't like it>look into different Kimi models>K3 is supposedly better than K2.5, and also shows up as having more than twice the t/$ value on ST despite the weighted averages on OR being higherWhat the fuck?
9
>>115924970Can Kimi do smut? I found GLM to be alright, and the Xiaomi models to be stupid.
>>115926600Yeah, K2.5 is completely uncensored (or at least that's what it feels like) and K3 doesn't seem any different in this regard so far.It's just weird how much cheaper K3 is right now, so I wonder if there's an invisible discount or something similar going on at the moment.The reason why I started looking for alternatives, other than the increasingly long wait times, is that K2.5 will sometimes take an eternity just to spit out an empty response, and still have the gall to charge you normally for it. Sometimes it works fast and flawlessly for hours, but other times it decides to hit you with this kind of bullshit for seemingly no reason.
>>115598760Life wife
Happy weenie my dudes.
Anthropic continues to enshittify things
>>115934490what kind of candy do bots eat
After a very long slowburn session, I have to say that Kimi K3 is not as good as Kimi K2.5.Near the end it turned repetitive and passive to a frustrating degree, but then I changed to K2.5 and everything went back to normal with a noticeable increase in quality.
>>115936961Kinda weird, you'd expect the newer model to be better. I guess they're just codemaxxing like everyone else?
>>115937315I don't know, maybe they did something weird behind the scenes which would explain why it's cheaper than K2.5, at least at the moment.
>>115936943Candy corn.
>>115937567Holy moly, it is expensive compared to DeepSeek and has less context. K3 is showing as more expensive on OpenRouter, though?
>>115939245I also thought that was weird, but ST says K3 gives you more tokens/$ than K2.5, and it didn't look like my credits were going down that fast when I used it, like 0.01 to 0.02 $ per reply.I don't know if it's something wrong on my end or if the difference wasn't noticeable when I was checking it or if I'm just going schizo.
