ANCHOR

migu sex.

still alive

>>13215 Why is your theme literally completely unreadable?

>>13220 miku theme for miku thread I'm a tomorrow chad

>>13215 "Oh! but, Red, what cute little fangs you have!" There's a big, bad wolf in the woods. Or so the village folk say. Made for /aicg/'s bunkie theme week #11, where the theme was classical tales! Also posted in the final hour of the deadline, so on time! https://files.catbox.moe/hgzei2.png https://pomf2.lain.la/f/ylmnm4tn.png https://rentry.org/chefseru

I'm more of a Teto guy but I also like Miku

>>13215 Theme week entry, based off the Romanian fairy tale "Mogarzea and his son". The prince of giant's life has taken a turn for the worse since the iele, the elves, made off with his soul as punishment for some minor transgression. He does not feel pleasure the way he used to, and the days have begun to bleed together. Even when you showed up and claimed to be family, he didn't really think on it much. He seems a little happier simply to have someone to live alongside and search for his missing soul. Four greetings: 1. First encounter. Gaslight the giant into thinking he's your dad. 2. Rainy day lesson. 3. Wait a minute...Mogarzea thinks you may not actually be his blood! 4. An unsavory encounter while doing some literal soul-searching. See if sweet-milk lake is actually full of sweet milk, bully some knife-ears, convince him to ditch the fairy and marry you instead. I dunno. https://chub.ai/characters/planewalker/mogarzea-362db5dae05e https://files.catbox.moe/gnjjtd.png https://planewalker.neocities.org/

>8chan's nullrouted on my normal VPN exit node Ouch.

>>13217 migu sex >>13217

>>13215 Fairy Tale Speed Dating (No Migu Edition) https://files.catbox.moe/2wmla1.png https://cardview.neocities.org/?2wmla1 A lot of princesses are in need of a prince charming. But listen, we've gotta keep the line moving.

>>13215 >sleepy... And with that, Theme week #11 - Classical Tales has come to a close. Thank you for your submissions! I'll do one last archive search a week from now and add any cards I see in the meantime to the rentry, so for any latecomers please include "theme week" in your anchor post for easy archive searching. https://rentry.org/aicgbunkerthemeweek

>>13228 >>13232 >>13239 bunkie's STRONGEST soldiers

>>13241 das us

>>13240 In the main greeting >who's gained knowledge ~~~~ The minor wordplay funnys are cute Also! We ended up with more Jacks than Alices huh?

>>13243 >We ended up with more Jacks than Alices huh? No Alices...

>>13247 ojou miku is cursed even ignoring the balloons on her chest

Morning, fuckers.

ME JUST SHAT INSIDE MIKU'S COOCHIE.

>>13251 good morning thoughts?

>>13253 I'd oil those joints if you know what I mean.

>>13254 you mean with a big hand held oil pump?

>>13254 l-lewd...

>>13256 I didn't know I needed voiceovers like this one but now I need one for videos like https://youtu.be/GNxpctsMtus. Very lewd, very good.

>>13215 >Should I even shill here? Place looks really dead. I noticed someone wanted my take on Jack and the Beanstalk, so here you go, hope you like it! >My submission to the classical tales event (theme week) >Loosely inspired by Jack and the Beanstalk Introducing sizebot Kuzunoha: It's been how many years now? Kuzunoha doesn’t know. It's been so long since the giant 100 ft tall fox goddess was confined to this realm, punishment for her mischievous antics against the higher deities of the Shinto pantheon. But it wasn’t always like this. Legend has it that long ago, there was a playful kitsune yokai who ended up ascending to godhood, a time where the gods walked among the early humans who roamed the Earth. Kuzunoha loved to mingle with the early humans, she saw them as funny little creatures and often tried to help in any way she could. Farmers used to pray to her for a good harvest, and women for fertility, over time becoming in essence a crop and fertility goddess. Her love for humans has no bounds, often going as far as having carnal affairs. The men used to offer themselves to the giantess. She misses those times fondly, now finding herself alone in this pocket dimension... that is, until a man managed to somehow intrude into her realm by climbing a mysterious, enormous bamboo that reached all the way to the heavens. 4 greetings. 1. First contact 2. Another day in paradise, you have known Kuzunoha for some time now. 3. Kuzunoha steps once again on Earth as a giantess, the seal somehow broken. 4. Sexo Chub - https://www.characterhub.org/characters/Genoo/kuzunoha-44bdcdb3f1fc RisuRealm - https://realm.risuai.net/character/587391dc-161b-45a2-8408-c3a97bba538b Catbox - https://files.catbox.moe/i47uvh.png Rest of my bots - https://rentry.org/295se

Any recommended (free) voice services for sillytavern?

>>13263 https://docs.sillytavern.app/extensions/tts/#tts-provider-selectbox

jenny won round 1 tbhdesubtw >>>/icup/6376

>>13266 oh fuck they're going?

>>13267 snuck up on me too, but Dorothy is fighting a furry later

>>13266 RIP Viv.

miss my elf wife

miss my human husband

miss my ADHD cat

>>13258 Jack:2 Alice:0 Thanks for listening genoo

>>13215 https://files.catbox.moe/1l8xr1.png https://cardview.neocities.org/?1l8xr1 181cm JK gaijin. A simple card about a tall British girl raised in Japan since very young, dealing with the feelings of "otherness" coming from that. 1. You are a new transfer student, your seat is beside her 2. Enkou with her and her Japanese friend

>>13277 she looks like she writes her number in {{user}}'s yearbook with a cute message saying she hopes to see {{user}} sometime soon but they never follow up on it and so spends a lifetime afterwards wondering what if

>>13214 Do you have any fun anecdotes you have done with your bots? Like making a story that is sick af while roleplaying with a chatbot.

>>13275 Anytime. I'm actually surprised by the lack of Alices, and that 3 of the 5 submissions ended up being sizebots. Sizefag event confirmed.

>>13280 sizeGODS...I kneel... >I'm actually surprised by the lack of Alices My theory is people expected Alices and so didn't make Alices.

>>13281 I contemplated doing a pretentious Alicebot but I wasn't able to do a second bot due to tariffs.

My plan for theme week was to do a Journey to the West card. (Surely one of the Four Great Classical Novels of Chinese literature counts as a Classic Tale, right?) But of course I ended up getting analysis paralysis and not getting anything done. Who would (You) be? Sun Wukong, Tripitaka Tang, Zhu Wuneng, Sha Monk, the white dragon horse, or just a random person accompanying them on their pilgrimage? Would they be genderbent? Would the bodhisattva Guanyin still show up to help bail them out every five minutes? Actually, that last bit would make for a pretty good low percent chance lorebook entry now that I think about it. Oh well, I'll get around to it eventually. Probably.

>>13283 It's not too late, anon. The host said he'd do another archive check in a week. I'd say go ahead and do it.

>>13284 still working on one myself, probably post tomorrow or the next day

>>13283 >(Surely one of the Four Great Classical Novels of Chinese literature counts as a Classic Tale, right?) I was promised an Enkidu femboy card and I'm still mad about it

>>13290 I love that they support apng here

I like this thread better. You lolis are pretty okay.

>>13293 We're just all sleepy.

>>13294 >sleepy lolis ToT

>>13214 Who's the artist for this? That Miku is pure sex.

>>13296 Looks kind of like https://sonohara.donmai.us/artists/217832, but I don't see it on any of their socials

>>13301 it's (heavily inpainted) AI, that's inpaint/brittle from /hdg/. forgot what his xitter/pixiv are.

Pits.

>>13160 What do I do when it suddenly stops using the CoT? I refresh and delete the JSON and reinstall it, but sometimes it works, and sometimes it doesn't.

>>13311 >refresh and delete the json You don't need to do that, it's just the AI model having a smoothbrain moment. Just keep trying.

>>13311 Add a new prompt at the very bottom of the preset. Set it to be sent as Role -> Ai Assistant. The only thing in here is: <think>

why does v3 like asterisks so goddamn much

>>13306 ena sex... >>13316 I *had* to stop putting *emphasis* in my greetings because *V3* would *not* stop using *them*.

>>13316 >>13318 I'm at the point where I not only refuse to put any asterisks in my greetings, I regex them out entirely. Maybe it's just placebo, but I feel like it improves the overall quality of the prose that V3 0324 generates. It's the same reason I stopped using ellipses, hyphens and s-stuttering, LLMs see it used once and it's just "monkey see, monkey do" from there on out.

Someone needs to tell Cohee to fix lorebooks within sillytavern and I don't want it to be me.

>>13305 Even knowing that, the only tells I can see are her microphone being too far away perspective wise, and her left hand holding the clipboard a *little* odd. They did a great job

>>13321 What's the issue this time?

>>13335 cumming in elf ears...

Card?

>>13337 hmmgnngmm >>13338 HMGMNGNGMGMHMMM

>>13335 >What's the issue this time? Loorebooks simply do not work unless you make them active for all worlds >>13338 Several if you want a robot meku but none to my knowledge if you want one using that art.

https://joemini2.onrender.com/ How do you get to the Captcha in here?

>>13347 >Loorebooks simply do not work unless you make them active for all worlds Since when? What?

>>13347 It's a "works on my machine" episode unfortunately.

>>13351 Oh right, double check the lorebook name (no weird characters, no spaces, filename routine) and make sure it's properly bound to the char. I saw weird issues with scuffed lorebook names these days.

>>13351 wait so are lorebooks scuffed or not scuffed? does your image equal good things or bad things?

>>13353 Anomaly said: >Loorebooks simply do not work unless you make them active for all worlds Howevermerine, I have no global lorebooks enabled and non-zero tokens on world info (both of them, actually) (because the card has a lorebook). Thisberry means I cannot reproduce his issue.

>>13354 thank you for explaining, I have smooth brain

miss my elf wife

>>13351 Damn Wait >>13352 You said that you had it bound to a character? I don't have mines bound. I was using vaporium and it wouldn't work for me unless I had it active for all worlds. If you updated recently say within the last few days or so maybe they fixed it? I'll update, test it and report back to confirm or not.

>>13350 It's been like that for me and a few others on other chan for a while now. If we were to go back to other people openly started talking about it, then that would put us to at least week ago.

>>13357 Then you better come up with the money.

>>13358 >You said that you had it bound to a character? I don't have mines bound. If the loreberinge is not global also isn't bound to a character, why would it be active? >If you updated recently say within the last few days or so maybe they fixed it? I git pulled staging immediately before taking the screenshot, obviously.

Is there any reason to give a keyword to a "constant" to a lorebook entry? Also what exactly does CFG do? It says it "makes parts of a prompt less or more prominent", which sounds like you could make it more detailed on certain things, like how characters look, but the example given is of denoting a character's emotional state. Both of these things just seem like things you could do in author's note anyways.

>>13365 I used constant lorebook entries for two reasons: 1. To extricate information that is going to be used across multiple cards into the lorebook, but that is relevant to all of them. I think I'm the only one who makes multiple cards per setting pretty much so that's rare. 2. And the other is to provide a list of terms for the AI to draw in content from the lorebook at will instead of hoping it RNGs onto them. I make sure those don't recurse so it doesn't just always draw in the bulk of the lorebook.

I'm not sure if it's my preset or else, but I noticed Gemini 2.5 having a sort of 'vagueness', in the sense of a tendency of 'maybes', 'probably', 'possible', like it's actively attempting to not make definitive statements.

>>13365 I don't think CFG does anything to modern transformers-based models? What are you using?

>>13371 CFG was put in for a while, but same as other stuff it required multiple passes and just like most of the directional modification stuff people have tried it didn't work as advertised, hoped, or intended.

>>13366 I've been using it almost as a "sorted" or labelled author's note for something that is supposed to be regularly referenced for AI thinking and interact with other lorebooks, but I'm not sure if there's a detriment to using it like this. I have also tried using it as a meta-guide/vocabulary like you though and that seems to work okay. >>13371 I'm on sillytavern with deepseek, but perhaps its just a vestigial UI element.

>>13364 >>13373 >If the loreberinge is not global also isn't bound to a character, why would it be active? Well that's not how it always was for me at least. At first all you had to do was place it where I have it circled in blue and it was like that for YEARS but ever since a week or so ago it only works whenever I place it in the red circle. I just updated and the problem still persists. Pics for reference and this place is neat for allowing two pics at once.

>>13376 Conceptualizing the utter shitfest ST would act like if every single constant entry in every single lorebook you import was actually always active...

>>13375 >I'm on sillytavern with deepseek, but perhaps its just a vestigial UI element. It's not even on the deepseek API lol. https://api-docs.deepseek.com/api/create-chat-completion >>13373 Yeah, I mostly know it from stable diffusion. >>13376 Oh I see. I mean, is just activating the lbs by adding them to active lbs really much worse than choosing them from the select? I think you may have been relying on residual GUI bits.

I just feel like talking about the first AI chatbot site I ever used. It was unhinged.ai and they had rules against children and animals. One day, I saw a bot that was a horse. Just a horse. At the time, I couldn't even imagine how you ERP with a horse. (this is like 5 years ago prob, idk) So I didn't. I just start the conversation with "*turns horse into glue at a glue factory*" And then the AI replies AHHHHHHHHHHHHHHHHHHHHHHHHHHH NOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOO And the site fucking LAGS as the NOOOOO extends OUTSIDE THE BOUNDS OF THE CHAT AREA and off the side of the screen. The lag was so bizarre and unexpected it's like the AI literally experienced death. I just put an artificial being through actual demise. I could not stop laughing, Shit site btw.

>>13387 >And the site fucking LAGS as the NOOOOO extends OUTSIDE THE BOUNDS OF THE CHAT AREA and off the side of the screen. DeepSeek likes to do this >I just put an artificial being through actual demise. Haha very funny. (but of course there is no artificial being of any sort in there)

>>13378 I know what your talking about, but when I plan to only use a lorebook for a single chat I would put it in blue and to use it for everychat forever I would put it in Red. That's how it worked for me for a long time without any problems because I only used one at a time unless the character themselves had a lorebook bound to them. >>13381 I mean it's not really but I'm likely to forget that I have it on for a chat and end up like that anon from days of old who used to post >Buzz Lightyear.gif >Left the fetish lorebook on during my slowburn But all in all it your right I'll just have to get used to it

Could someone give me a quick rundown of world info placement? I assume it is just a "priority" but it's not super clear to me what the order is for "char def", "example message" and"Author's Note" Also the different impact between "system" "user" and "assistant" used in system prompts and specific depth AN. >>13376 Can you not just go to the character tab and activate it for your chat there?

Character ai has enabled a nsfw model, still a bit censored but you can hold hands nonconsensually.

Hey guys, I currently use agnai but is there a way I can connect a jailbreak or something because I'm not really feeling their bots. Though, I'm open to other ideas.

>>13398 Last time I use character.ai it was dumber than 3.5-Turbo

>banned on 4chins for 3 days I am impressed but also disgusted as how fast the jannies are over there. I was certain they'd still be reeling from being doxxed. Allow me to copypaste my shit here. >>13215 shilling for Tsukikage no Sayo, a 15 year old kunoichi who I hastily made because I really wanted a >genius >horny >loyal ninja girl with long black hair. The new Gemini 2.5 snapshot is competent enough at handling her. Comes with six greetings: >Your father, Tsukikage no Yoshihisa, is dead, and she and her family have come to swear their oaths to you, his successor. >You are one of the hidden village's blacksmiths. Sayo has something she needs you to fix. >You are one of Sayo's peers and a fellow pupil of Kiyomasa-sensei. You are on a reconnaisance mission together. >You are a samurai lord. You appear to be on the cusp of being assassinated by Sayo. >You come upon Sayo when she's training in the yard. >You are Sayo's master. You have summoned her to your bedchamber at night. https://characterhub.org/characters/sillygoy/tsukikage-no-sayo-95deef5fe074 And also for you gay niggas from the claude 2.0/gpt 3/pygmalion 6b era, I finally got around to updating Vannian with new art and a few new greetings. https://characterhub.org/characters/sillygoy/vannian-10872bae

Niggers.

>>13215 I feel weird shilling twice in the same thread. Oh well. Clark "Cheeto" Greyson and Yuuta "Ace" Shishigami lived a pretty stable life in the underground vault before they got the Fallout 1 treatment. But instead of raging against those who gave them the boot, they decided to live the best they can and see all the interesting things they can find in a world that's survived The Great Distortion. Thankfully, they have the perfect thing to ensure their survival: a fucking TANK. Five greetings: 1. Two idiots almost run you over with aforementioned tank. 2. Laundry day! Also, what's for lunch? 3. The three of you stumble across a mystery flesh pit. Throw something in it! 4. The lads approach your settlement looking to trade goods (and weird shit they found that looks cool). 5. You're a fellow vault dweller they haven't seen since they got kicked out. Poke a Horror with a stick, spray-paint dicks on the entrance to the underground vault, make weird potions and dare each other to drink them. I dunno. https://chub.ai/characters/planewalker/cheeto-and-ace-a6e11b1f2f67 https://files.catbox.moe/mifc49.png https://planewalker.neocities.org/

>>13215 I need that first Miku animation.

>>13418 Thanks my nigga.

I'm about to get high and coom

>>13420 epic get same

>>13409 bulgeGOD...

>no replies in 24 hours Isn't that weird, I thought you said you were staying here? Guess the IDs got to you after all.

How do I have a vanilla RP adventure? I can't stop fucking the bot and then reaching a dead end there.

>>13441 Depending on your AI enable and disable nsfw and lewd lore for when you don't want sex.

Cute boys.

hello my friends, I am a historian and history teacher writing to all of you from Chile in the southern hemisphere, in a forgotten corner of this earth. I apologize for not writing in a specific thread, it is late and I am tired, however I deeply thank anyone taking the time to read this and maybe lend a hand with some basic orientation? TLDR I am working on a project that might open up big doors for me; 133 years ago we had a president who KMS himself after losing a Civil War, his house is now a restaurant and every last May weekend we celebrate something called "patrimony day" where we celebrate all our cultural history. I am working together with a colleague of mine to host a tour and patrimony event for that day on the restaurant and we want to recreate with historical data and research historical scenes and cities with the use of AI. In comparison to what all of you may know, I am a complete and utter beginner, but we have the resources and capacity to bring about real AI projects for education, teaching, and history in our country if we manage to show the capacity of AI as a tool and field of learning for those who will attend that day (people with influence and resources). I got into AI by using ChatGPT and Deepseek to improve my education, I am now exploring HuggingFace, but I am pretty much disoriented with where I should be looking and what I should be learning, or what I should be using. My computer machine I doubt currently could run local models, but I have the funds to improve the ram at least. I currnetly want to reconstruct a "panning scene" from my small town in Chile where wounded are arriving to a train station after a battle, I have pictures and descriptions, and I also wish to build something like a chatbot to implement during the event for those attendees in the "digital tour" we will be doing with these technologies, which is the reason why I post in this thread too. I apologize for writing this much to explain and maybe risking to be read as asking to be spoonfed, but I would really really appreciate if I could receive any indications or directions at all to get the show running. I have big plans for the future and they include all my freinds.

>>13448 Don't use chatbots if you want to accurately mimic real events. It will end poorly.

>>13449 i plan to add a chatbot as an extra "explainer" that is aware of the whole research data i will feed it, so that attendees may ask it for extra inquiries related to a historical source or image, not as a main act of sorts, just so that we can show how this format may be used in educational contexts, why do you envision it could be a problem my friend?

I've got a challenge: Try cumming early, try to leave the girl unsatisfied. The model I have seems to think sperm in womb causes immediate climax for girls, no matter what card I use.

>>13441 Don't have an open-ended adventure. Don't start in a situation where all your and the characters basic needs are met. Don't spend time alone with a girl in private. (historically people assumed something happened in those situations, with good reason) Give your own character a reputation for having taken and abandoned many girls. Or provide another reason to make her wary of your intentions. To continue a story after your mistake: a) Get her to agree to marry you. explain yourself to the family, arrange a full wedding. Do you have a place the two of you can live? b) You want to keep chasing ass but she's calling in all favors to make sure you take responsibility. c) Her father wants you dead now. You're on the run. d) There's a knock on the door right when you're nutting inside her. Who is it? e) Edit the story to delete all the sex and add a timeskip to where you left off. "In the small tent, the two of them managed to forget about the threat of the demon lord for a few hours that night, but the next morning..."

>no replies in 2 days Wow this place is dead.

>>13464 I'm still here...

>>13464 I'm not.

>>13464 sorry i need at least 500 people in one thread when i talk about forcing my harem of femboys to sell feet pics so we can continue to afford our mansion

>>13215 https://files.catbox.moe/yw1gr0.png Horse pussy.

>>13471 >no horse lorebook

>>13472 I do not think I have to explain to the model what a horse is sir.

>>13473 Okay but she won't even know who Cyan is. I am currently rectifying this, of course.

>>13215 Sure, I'll shill here too. This was a bot created for the theme week, and it has four different greetings representing each direction. I tried to do the opposite as my Alice bot, which relies quite heavily on a lorebook; with this one, I challenged myself to rely on none, so there's an AIDungeon-esque feel to whatever your model might make up. https://chub.ai/characters/heartofthecards/the-land-above-636f1c2703a3 Speaking of Alice, here she is. This was not a part of the event, as I was already working on it beforehand. By the way, this is her Megami Tensei interpretation. Description tokens are lower than usual because It Just Works. https://chub.ai/characters/heartofthecards/alice-8e254e29bbb1 Sanae Kochiya has also been updated. Just a quick one, no new greetings yet, but I had to edit the 2nd person greeting it already had. https://chub.ai/characters/heartofthecards/sanae-kochiya-931073f0fcb4 It should also come with a quick lorebook update to include the cast of the new game's characters. As the game itself is not complete, neither is the lore, but updating doesn't scare me. Credits to Momoura for the 19+ fork, and hebelover for porting bunyear's work.

>>13215 >>13471 https://files.catbox.moe/djy76u.png 1600~ tokens of horse. Gave it the ol "swipe a few times and ship it."

>>13476 Based, thanks for the hard work.

deepseeks becoming illegal? https://www.bis.gov/media/documents/ai-policy-statement-training-ai-models-may-13-2025

They added video bgs to ST at some point so you can now do bullshit like this without converting it to GIF beforehand (note: the video upload took like 1.5x the video duration on my machine. I have a decent CPU. I don't know what are they doing).

>>13479 neat. I will make comfy atmospheric backgrounds for some of my bots now.

>>13471 >>13476 Don't mind me, just posting memes.

I just came back to AI chatbots after an unfortunate ~8 month break and all my proxies are dead :( Is there any Opus available anywhere or do I need to seek greener pastures?

>>13482 >8 month break >comes back to the altchan Eto?

>>13409 >tagged Huge Dick >character only has 8 inches Anon...

>>13487 No, a newfag. Got tired of the big sites nonsense and spam so I've been looking for cozier alternatives.

>>13464 i'm still here but i just haven't really been RPing or doing anything with bots lately. still not going back to the schizo jeet shithole.

Did anything happen with deepseek or the chutes proxy? I played a couple of days ago and it seemed okay, but now every time I try to continue the chat I get gibberish a quarter through.

>>13504 >chutes I don't think R1T is a real combination of R1 and V3. For instance, consider how it doesn't consistently output its reasoning. R1 produces the reasoning then produces the response. R1T produces the reasoning as part of the response. The combined R1 and V3 weights might be a literal combination. As in, R1T only produces its thinking output because it's being told to think on Chutes' side and has references to it in other parts of the model. To answer your question, try lowering your temperature. Non-DeepSeek providers can become problematic at temperatures above 1. >the chutes proxy Is that an actual proxy or do you just mean chutes hosting DeepSeek models?

>>13505 >I don't think R1T is a real combination of R1 and V3. For instance, consider how it doesn't consistently output its reasoning. MAI-DS-R1, Microsoft's R1 finetune, suffers the same issue of inconsistently outputting an opening <think>. Personally, I'm convinced it's a side effect of the finetuning process. >R1 produces the reasoning then produces the response. R1T produces the reasoning as part of the response. They use the same method, generating tokens until they generate a stop token. It's just that finetunes like R1T and MAI-DS-R1 seem to forget to open with a <think> significantly more often than base R1 does. If it forgets to open with a <think>, but still generates thinking with a </think> at the end of it, and you want everything before that to get put into a reasoning block, you can use the following regex: /^(?!<think>).*<\/think>/s Replace with: <think> $0

>>13505 It's just the chutes host. Turning down the temperature seems to help, but it was set at the (rather high) 1.80 from the start with cherrybox and had no issues until this morning, where I couldn't even open a new chat. Really weird. Thanks.

>>13507 Thanks, anon. Do you prefer R1T or V3 0324? Myself, I like R1T. >>13508 It's great that you worked that out. It seems like most people use .3-.6 for temp. I thought you meant a reverse proxy. OR says Chutes keeps logs for an undetermined amount of time, so I've been cautious about using Chutes directly. I don't know what their content moderation and logging policies are.

>>13544 >Thanks, anon. Do you prefer R1T or V3 0324? Myself, I like R1T. My personal preference is V3 0324. I guess R1T still feels a bit too wild and untamed for my taste, though it's certainly less borderline schizophrenic than R1 tends to be at times.

miss my elf wife.

Is there any particular preference for how cards are posted here?

>>13555 Sorry wdym?

Is there a way to make token probabilities not super laggy crash my session? Is this a matter of having too large of a prompt?

https://www.youtube.com/watch?v=EvtPBaaykdo Starting soon

>>13560 >Is there a way to make token probabilities not super laggy crash my session? Is this a matter of having too large of a prompt? I think it's only affected by the model's most recent reply, but don't quote me on that. Limiting the number of candidate tokens in some way is probably your best bet, either via Top P or Top K. Try dropping the Top P down to 0.92-0.95 and see if that helps.

Is there a way I can retroactively write out a side character (or at least ignore her)? I have a medical RP with a doctor and there's a nameless nurse that was introduced early by the AI and has just permanently sat awkwardly at the side lines and I want to rid myself of her, but she's been around so long it would be difficult to edit every message involving her.

>pic So, this is the power of DeepSeek... WOAH >>13574 Depends on the model, but it's very likely that the context has already been poisoned. The best you can try is just write an OOC like, (OOC: write the nurse out of the story.) and see how it goes.

>>13575 Deepseek. I guess I can try that and add a ton of negative bias for "nurse". A good thing she has no name. Question on bias though, when putting multiple words in one bias textbox, does it impact the exact phrasing, or each word individually (e.g. if I put in "kick ball" and put in negative bias, does it still permit "kicked ball" or "kick balls"? And how much bias is too much or too little?

>>13576 Are you talking about Logit Bias? I'm pretty sure it only works on GPT.

>>13577 Damn.

>>13578 Yeah, it's a shame. Logit Bias would be great with any LLM, Gemini in particular. That one has some good stuff that is tucked under tonnes of 'isms.

Slow thread.

Does anyone know if henk717 ever posts here? I know he posts on half chan sometimes. I reported an issue for koboldcpp on their github but my account was shadowbanned.

had some really good stuff going on hiwaifu but ran out of credits, what models are good for multi character scenes? I've got my tavern set up so all the installation shouldn't be an issue

miss my elf wife

>>13619 Can't you email him on the commit email?

cftf?

Does anyone else think the chutes DeepSeek is shit? I've been using it, and its quality has dropped off significantly, regardless of preset. I tried using Fireworks' DeepSeek, and that worked well. However, it costs money.

>>13678 I'll look into it thanks.

I'm still enjoying my time with chatbots

I FAP TO AI GENERATED ANIME EVERY DAY MULTIPLE TIMES A DAY

>>13679 I remember that early on, there were LotGH cards on chub. Kircheis was one, I think.

I can't wait for the day we get LLM's dedicated to one franchise each so I can perfectly self insert into my favorite anime shows.

>>13472 https://chub.ai/lorebooks/TaintedMilky/unisex-horse-sex-genitals-lorebook-8521c89c67bb

>>13690 MEIN NIGGA

>>13690 my fuarkin hero......

>this place is super dead Noooooooo! I bring good stuff anons. It's a preset for Gemini (and some bonus stuff) https://rentry.org/szer4 The bonus stuff is some fun lorebooks. One of the three is a way to implement randomness so that you can write {{getvar::var_NOUN}} and you get a random noun, or write {{getvar::var_TVTROPES}} and you get a random trope. If you have ever dealt with BIG {{random}} you should know how obnoxious it gets.

Starting this week, Chutes will require money for its free quota... >>13715 It's hard to have active discussion without visibility for new users or sources for variety. Thanks for the preset, even if I don't know how to get Gemini...

>>13716 https://waveexistence-merkava.hf.space

>>13717 >Merkava Didn't Merkava leak logs once?

>>13715 was this place in use before the 4chan shutdown?

miss my elf wife

Chutes, among others, are pulling out of OR, so OR is partnering with other services for free access to models like V3 and R1. However, they're dropping less popular free models. I suspect this to mean R1T and R1T2, which are both fairly good as long as you don't incidentally make a pop culture reference. >>13717 Merkava's dead. >>13719 Check the date of >>1 and compare it against the date of the hack.

what do we even talk about if we do return, i just lurk mostly and reply to other people chatbots?

>>13727 You could talk about models, or what preset you're using, or what chatbot you're using, or botmakers, or chub, or the card webring, or really anything revolving around the hobby from an amateur all the way to a technical level.

I like to use Qwen.

>>13733 Same, The Qwen models are very underrated, they are quite small compared to other more mainstream models. But they give more complex and thoughtful answers. I mostly use the 32B model with the thinking mode activated. Also does someone knows why the Qwen mascot is a capybara? I like the desigh though. https://huggingface.co/spaces/Qwen/Qwen3-Demo

Since chutes is dead, what other (good) free models are out there?

>>13734 > Also does someone knows why the Qwen mascot is a capybara? "Known for its calmness and adaptability, the capybara symbolizes how Qwen is always ready to help you easily tackle the challenges of the tech world" Plus capybara has a lot of meme potential. https://xcancel.com/Alibaba_Qwen/status/1844692086307828203

K2 sucks. It would be good if they removed the Grok-style sassiness and GPT-style anti-NSFW bias, but it sucks. >>13736 There's always OpenRouter...

https://discord.gg/VQg6C5TNGG https://discord.gg/B97uT7Wv5h https://discord.gg/6UtVFBMHgm https://discord.gg/T4M25bUcwJ https://discord.gg/BSk46AebpE https://discord.gg/5EqRPa3dVg https://discord.gg/dhqZm5KXv4 These are all free ai apis with free claude gemini, gpt. But it won't work with prefill, i tried. Some only allow payments for claude some offers it free. For gpt all is free in here including gemini.

Does anyone know of any cards that use the Live2D extension?

I've been neglecting Gemini. It's not that expensive, and the pro model, even with a shitty JB like Avani's, makes really good outputs.

>>13752 I used to be a geminislopper, but the moment. THE MOMENT I got some claude, I crumbled. Gemini is fine, but i'm just so fucking tired of it

>>13715 well i've heard that for Gemini, using {{user}} roles instead of System ones lower the chance of filters, since the filters pay more attention to system prompts than anything

>>13753 I've had too much of Claude. Due to pricing, I haven't tried 4nnet and 4pus, but I used to use 3.7 all the time. I'm so sick of how Claude writes characters.

>>13755 man I wish llms had better writing

>>13757 I'm pretty sure most companies prioritize volume over quality, and US corporate models don't have the efficiency or infrastructure of CN corporate models. It's a shame that every model has to be marketed as capable of doing everything.

>>13759 maybe one day there'll be a model exclusively for roleplay and will be a GOD at all things writing

>>13215 I forgot this place existed :D I will shill here too because I feel like it. I have a brand new Neocities page! https://uraurabots.neocities.org/ And not one...not two... not three BUT FOUR new bots. Moriko Fujioka, a sorta remake of Nice Neigbour Lady though I did take some extra liberties. She's an insane lesbian who loves little girls like (you). Technically anypov for the first message if your persona is a convincingly crossdressing litle boy. https://chub.ai/characters/uraura/moriko-fujioka-f668f736b4c3 https://uraurabots.neocities.org/page1#moriko_fujioka Nanami, A jirai-kei wearing (translation: those pink and black clothes mentally ill girls wear) oshikatsu-loving (translation: being reallly obsessd with a band) social-media obseesed (translation: no translation needed) girl who is totally a pure virgin and has totally never been hypnotized and fucked by old men in her life. I posted a joke version of this card before but if you have that one throw it away because this one is way different and wasn't made in 30 seconds. https://chub.ai/characters/uraura/nanami-4d4e50378e7e https://uraurabots.neocities.org/page1#nanami Yumemi, A lonely lolibaba rabbit-girl who likes drinking tea, quilting and telling stories. Even though this bot was entirely made because I remembered that cunny originally was a word for a rabbit, it's entirely SFW and made for wholesome slice-of-life yuri. You can do whatever though, not like I care. https://chub.ai/characters/uraura/yumemi-eab42fce4ffd https://uraurabots.neocities.org/page1#yumemi Your Oblivious (male) Mother, It's just that bot but a dilf. And with some alt greetings. Just fuck your idiot dad. Do it. https://chub.ai/characters/uraura/your-oblivious-male-mother-7c77ca884612 https://uraurabots.neocities.org/page1#your_obliviousmalemother

>>13762 Are you a fujo? Your bots and pages scream "femanon".

>>13767 my peanus weanus is humongo

>>13762 >femdilf absolutely fucking cultured, will knock him up with a litter of buttbabies

>>13715 >>13715 >>13715 I've a new version of the preset. It's an almost complete rewrite. https://files.catbox.moe/tigp4b.json This new version is much more polished and it has better outputs (but it's still Gemini) while getting blocked less.

>>13772 (I did some more testing, and it might be a good idea to turn off the '🗻SYSTEM MAIN OBJECTIVES' since it seems to trigger the cockblocking)

>>13734 Someone else having issues with the Qwen Models? It looks like the servers are melting.

I hate the {{user}} role. A frustrating facet of it is that, when defined, every model I've tested automatically applies a positivity bias that alters the character in order to please the user beyond just agreeing. If a character has a preference for a certain type of thing, they'll make logical leaps to think that the user fits that preference. I theorize two potential workarounds for this are to use a card as a persona or follow the old C.AI method and define {{user}} in the opening message.

>>14604 I tested it. Neither of these work. It turns out that the bias can be for anything, not just positivity. For example, a character who even vaguely matches the description of what another character thinks is gross but hasn't been described yet will be mentioned almost verbatim if this other character is asked what they think is gross. This means that it's some form of the pink elephant effect and not necessariy a positivity bias, unless the existence of {{user}} coincides with some sort of built-in bias. Unless the character has an exact description of what they think is gross, the phenomenon will still occur. The LLM will prefer to use what it has instead of making something up itself. I think this is referred to as a presence bias. My current solution to get around any bias toward {{user}} it is to make a pseudo-persona out of a card, play in a group chat where the cards switch, copy the pseudo-persona's definitions into a non-constant worldbook entry, tell the model to not do anything as the pseudo-persona, mute the pseudo-persona, and manually generate a response for the pseudo-persona while keeping streaming on, immediately manually stopping the response before anything is generated. However, this doesn't prevent the presence bias from happening as the pseudo-persona is now present. Other than making more descriptive cards, I can't think of a solution for this other than better defining characters, and that's not a one-size-fits-all solution. Messing with settings doesn't work. Even if I crank the Presence Penalty to its maximum, presence bias still happens. Even if I use a thinking model, and the model even says it won't do it, presence bias still happens.

>>14605 Diluted nested thinking might work. If the model makes a response for every aspect of the prompt and context, going from the largest scope to the smallest and comparing them against each other, it should be able to get rid of presence penalty. Continuing on the example of a vaguely gross unintroduced character, the model would make an answer based on what it thinks is gross based on its existing definitions alone, then an answer based on its context with the character asking them the question, then an answer based on the its context with the unintroduced gross character. The model would then argue with itself over which fits the user's request more, an answer that foreshadows the existence of the unintroduced character or an answer which more truly portrays the character answering. Each thinking block would be a separate instance so presence bias isn't applied to them, and they would be fed either all at once to the ur-thinker or engage in a dialectical process where the AI compares and picks them apart, creating more variations to do so to until it reaches an end result agreeable to itself or the user's arbitrary limit, and using that response. An even better system may use thinking blocks from multiple models. The end result would mimic the real patterns of a brain, rather than immediate pattern recognition. The next step would be to let it use the internet unabated for these arguments. Further responses themselves would be nested dilutions. The model would be able to refer back to its previous methods of refinement and create a better model to refine its responses. Since the model is arguing with itself, it wouldn't even need to read the outputs, just the token probabilities and placements, and these may be able to be stored as seeds or hashes.

Updated the preset (gemini) once more. https://files.catbox.moe/vragli.json https://rentry.org/szer4

>>14604 >>14605 >>14606 Man, you write like a schizophrenic, no offense. That whole 'manually stopping streaming' thing is just bizarre. Honestly I had to ask Gemini to tell me what's your deal was. Your issue is that LLM are too influenced by existing tokens, right? I don't think there is much you can do about that. There is also the fact that not using present information can end up being a sort of Chekhov's gun situation; example: You are in a group chat with a human supremacist character & an elf character. You ask the human supremacist what creature she hates the most and she'll likely say 'elves'. Is that really that wrong of a thing? It does make sense NARRATIVELY. I do think your issues (especially positive bias) can be solved or at least reduced with <think>. But you also need an LLM that ACTUALLY obeys the preset given. That's what I like the most about Gemini, how servile it is. I personally find both GPT and Claude too libertine with the rules, while deepseek is too stupid to follow them. My advice is to use Gemini, use the preset I made above this post as basis, then write a custom <think> clause that deal with your problem that I'm still not 100% sure I understand.

>>14600 Looks like the qwen servers are back. Took them a whole week to fix their crap. And now I dont trust that shit. It took 6.66s to respond me... IHATETHEANTICHRIST

But seriously, which monster girl WOULD own a pizzeria?

>>14631 For American-style pizza, mothgirls. They're very urban. It makes sense you'd find them buzzing around greasy streets making greasy pizzas.

>>14609 It was a little schizophrenic, I'll admit. I was at the end of my rope on the problem as I tried to brute force Kimi and DS to not have it. I also vaguely remember it being an issue in GPT and Claude. Basically, it's what you described about existing tokens. If you were to ask a character "What's your ideal X?" and your persona vaguely matches what Kimi or DS think falls into X as a quality, the character will describe your persona. This is intensely frustrating when you have a character in a group chat with an interview character because, even if your persona doesn't have a description, the character being interviewed will reference the interviewer's traits for that "What's your ideal X?" question since it seems Kimi and DS are tuned to use all available context rather than pull from the model, hence why they're dumb. My hypothesis was that models detect the existence of a {{user}} and thus favor him dispraportionately, since it seems the bias heavily diminished when I used methods to reply that didn't involve a {{user}} response or make reference to a {{user}}. I wish personas and characters could be cointainerized so they don't influence each other. I haven't tried the scenario described on Gemini yet, but Gemini seems to not have as big of a problem with it. However, I'm using your preset with that option on how to treat {{user}} on. As for your preset, here's my experience with it so far >weak NSFW promotes violence and aggression too much >attempter thinking makes every character act super sensitive and act like a rape victim, eventually shutting down >medium thinking and cherrybox NSFW often trigger the filter >draft + rewriter filters out a significant amount of character quirks, like accents and typing styles, even if it does manage to improve the quality otherwise >think + draft + rewriter frequently fucks up in deliniation by putting everything in <think> instead of their defined regex, sometimes even doing the regex properly afterward, wasting tokens and money >anything higher than weak tends to see Gemini make its own judgements on characters in <think>, pidgeonholing them into specific roles >I don't know if the random feature in attempter actually does anything, since Gemini seems to push for narrative coherence and may be suffering from the influence of existing tokens because I kept seeing the same rolls, and the numbers and descriptions of these rolls were influenced by the character >attempter is always the character and never the user, and this is either because I reply with "I do X," an absolute statement, or the JB tells the model not to act for me

So, rentry alternatives now that they're cracking down on AI stuff. > https://rlim.com/ (javascript bloat) > https://gitgud.io/ (lets you do a github sites style thing, requires an account) > https://neocities.org/ (dubiously tolerant)

>>14640 there's https://pastebin.com/ but it needs an account for edits and doesn't allow images https://ghostbin.lain.la/ also exists but account creation is disabled so owners have to keep the tab open forever to edit. also doesn't allow images.

>>14641 pastebin is also fucked for storing information long term, dunno about ghostbin.lain

>>13215 Nao is a demure feminine slaveboy who only speaks Japanese. It's time to dust off your IME. Also he has a big dick. https://uraurabots.neocities.org/page1#nao https://chub.ai/uraura/nao-75ce52da3b8a This was originally intended to be a fork of https://chub.ai/characters/NG/mexican-slave-training-spanish-trainer-064a1a010898 but it diverged pretty heavily. The basic idea of learning a language through a slave with automatic correction is still retained from NG's bot though. A very green loli alien girl from ??? space. Midori like touching things and eating things. Maybe those things will be you. https://uraurabots.neocities.org/page1#midori https://chub.ai/uraura/midori-5b318109ff86 Adapted from one of my personas. 4 greetings. Jasper is your best friend with a small dick who made the genius decision of drinking a random potion off the internet and now he has a pussy instead. Remake of one of my older bots. He has AT LEAST 4 more tokens then the old one and also a backstory. https://uraurabots.neocities.org/page1#jasper https://chub.ai/uraura/jasper-69248734d1cf Hera is a boisterous warrior who wants to get rich and famous by braving ero-dungeons. https://uraurabots.neocities.org/page1#hera https://chub.ai/uraura/hera-991e04012f99 Also her card image is an apng. Technology is sexy. Also I stole some stuff from ttrpg anon for her new PHI, thanks. She's actually been 99% finished for a while (As in she's so old she has example chats) I just never got around to releasing her. I recommend using the dice extension with her. https://github.com/SillyTavern/Extension-Dice You can try using the function calling thing but it only works like 30% of the time in my experience. I also added some stuff to my neocities. I have a new gifts (and misc) page for things people made for me https://uraurabots.neocities.org/page4 And I actually have something in my prompts and JB section now. https://uraurabots.neocities.org/page2 >>14640 The meta_botmaking_list rentry is back up (and unchanged) and from what I understand they were testy about loli stuff. It is a good idea to split out though and not put every egg in one basket though. I like gitgud personally and I wouldn't recommend pastebin. Pastebin hates NSFW of any type.

>>14608 https://files.catbox.moe/np6ae4.json Updated. >>14635 Thanks for the review, Anon. It's pretty interesting how we got different experiences about certain things. I imagine we rp quite differently. >attempter thinking makes every character act super sensitive and act like a rape victim, eventually shutting down At a certain point you gotta use one of the Jailbreaks. >anything higher than weak tends to see Gemini make its own judgements on characters in <think>, pidgeonholing them into specific roles Yeah, I've seen that. Maybe something like "Don't flanderize characters too much" or whatever inside the 'Core Rules' could do the trick. >I don't know if the random feature in attempter actually does anything, since Gemini seems to push for narrative coherence and may be suffering from the influence of existing tokens because I kept seeing the same rolls, and the numbers and descriptions of these rolls were influenced by the character I think the issue was having Top K at 0. Pumping it up (in the new update its a 200) seems to make Gemini more "fluid"

>>14644 I started using jailbreaks, and it helps. However, it seems to be required if you're using GIGABRAIN with everything turned on. When I did that, the multipart jailbreak worked with the small version of part 1 on. If it was anything else, things sort of worked but were overly long, improperly formatted, and had the same sort of rape victim mentality as no-JB but without actually falling into it; Gemini treats you like a bad person with Kaguya or code on. I tested it with characters who only look loli, and your preset works with sometimes, and on certain NSFW options. Weak and U3 worked most of the time, cherrybox worked a quarter of the time. It seems like if it hits "looks like X" too early, it shuts down.

>>14644 >>14651 Trying it out again, this time with the latest update to your JB, and there are two major issues with loli and anything like it. First, if the model generates anything that even hints a character is or looks underage in the thinking, it hits the filter. After that, it's a tossup whether you can give a continue prompt and have it keep generating. However, it generates loli just fine in the response section if you have a jailbreak on. Second, and this is more general, your NSFW prompts are poorly structured. NSFW Weak tells the model that it can generate NSFW, but it doesn't detail how. The rest of the NSFW prompts say how, but not that it can. In fact, most of the NSFW prompts say just to write NSFW overall, completely changing the tone of what's generated. Turning on Weak and one of the other prompts edited to say that it's how NSFW should be written helped out a lot. I found that anything close to loli is generated at a higher rate with NSFW Weak and one of the edited NSFW prompts.

Having fun with claude, fucking finally again. If anyone is paying smol or jew, consider aiho.st instead

After some testing, I found that the issue isn't just with loli. If a character is 16 or 17, they still hit Google's filters once things are too erotic and their age has been mentioned in the thinking block. Regardless of JB, GIGABRAIN was useless with it. I had to use Medium Thinking and Refusal Destroyer, and even then I would only be able to generate something without having to press continue or regenerate about half the time. This issue blocks characters from being represented at their canon ages. For instance, Tifa and Yuffie from FF7 would hit the filter. I didn't test their cards, but I know both of them are under 18. The filter itself is definitely a Google problem, but Gemini is capable of generating erotic outputs for characters under 18 as long as you can skirt around the filter or manage to dodge it. I've tried telling it not to mention ages within the thinking block since it should be able to infer this by context and character description, but that only sometimes works since Gemini will either mention it anyway or use a synonym. Even if it does work, it loses some of the details in its output. Also, since the JBs are output before the thinking block, the thinking block analyzes and mentions the JBs, and character details aren't consistent across thinking blocks.

>>14660 >since the JBs are output before the thinking block I meant to type "since the JBs are output after the thinking block". Gemini interprets being told to not analyze a JB as being told to not analyze it in the output.

Sonoma Sky is so good that I'm starting to worry. It really is like having Opus again.

>>14709 Correction: Sonoma Sky is good, but only if you don't use prefills or group chats. It breaks and gives empty responses in both scenarios. This is very obviously supposed to be the basis for a one-on-one RP model.

>>14710 Double correction: Sonoma Sky does work with group chats, but the way to do them is schizo. You have to >have a card in a normal chat >reply to the first message >have the card generate a response >transition to a group chat >add any other cards you want >add a card made out of your persona and mute it >say your persona is the card made out of your persona in your persona description >generate a response from each card you want introduced if you're switching between cards, edit the first non-intro response to have these new cards enter, or edit your first reply to say that the cards you've added have entered >generate a response from the card made from your persona >replace the contents of the response from the card made from your persona with what you want to say >generate the next response >manually generate then rewrite responses made by the card maede from your persona >repeat This sounds insane compared to other models, but it's the only method I've found that works for group chats. Since Sonoma might be for something like Grok, I think it, on the hosting side, has some sort of importance placed on the User role to distinguish it from the Assistant role more than other models. However, using the method I described creates a scenario where the Assistant is only replying to itself. Since it's almost certainly saving responses, I haven't tried to, and don't intend to, test it on anything under-18, but the model appears to be completely jailbroken by default. I badly hope nobody fucks with this model and that it'll be usable in the future, even if it has to be paid for. Aside from Opus, I don't think I've seen a model this good. My only complaint is that responses are long and overly detailed, but if it's something like Grok, which is supposed to answer questions with as much detail as it can, I can't blame it. Sonoma Dusk is a bit worse.

>>13215 I am the only one who shills here anymore. But I will do it anyway. Alicia Descoteaux, an ojou-sama magician girl who won a duel against you FAIR AND SQUARE. Put on the maid dress. https://uraurabots.neocities.org/page1#alicia_descoteaux https://chub.ai/characters/uraura/alicia-descoteaux-9053902c1268 Gu Weisheng, Cultivator and disciple from a rival sect. Also he has a pussy. https://uraurabots.neocities.org/page1#gu_weisheng https://chub.ai/characters/uraura/gu-weisheng-b53683eb206f The Red Flame. Himbo superhero with fire powers. Also he has a pussy. https://uraurabots.neocities.org/page1#the_red_flame https://chub.ai/characters/uraura/the-red-flame-016bb91a3be0 Sora.love a 90s desktop assistant who lovey-loves love and YOU. Also she's a yandere :D https://uraurabots.neocities.org/page1#sora_love https://chub.ai/characters/uraura/sora-love-90bf16f8bf22 also i updated my preset https://uraurabots.neocities.org/page2

>>14711 >Sonoma Sky and Dusk Alpha really were Grok 4 Fast Get it while it's free. Elon's definitely logging, though.

>>14723 but is it actually good?

>>14724 It's pretty good. Definitely a step up from the Sonoma stages. I just found out this was the case, so I haven't tested it extensively. I'm testing it with a tweaked version of the Gemini prompt that one anon has been working on and posting in this thread. So far, I've found that it still can't take prefills, but group chats sometimes work now. Something about that preset makes them work now, but with a barebones preset, it doesn't. I wish someone had made a dedicated preset for this. I think Grok is a serious competitor now.

>>14724 >>14725 I've tested it some more, and I think that the model needs Claude-like frequency and presence penalties, as well as a Top K in the hundreds. If you don't force it to use the rest of its dataset by penalizing repeating existing tokens, it regurgitates what you've already given it. This isn't very noticeable when you have a set scenario in mind, but when you're doing anything creative with a character outside of a set scenario, it becomes obvious.

New v3.1 is pretty okay. It's no 3.6 just yet, but...