>>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...

I haven't used ST in months, are free models still easy to access and good, or are we still in a drought?

>>14735 if you want claude, then no if you want gpt, you can find it if you look around if you want gemini, deepseek, kimi, all the other small chinese models then you should be good theres also a new grok that some anons have been liking https://openrouter.ai/x-ai/grok-4-fast:free

Since this is a long-running thread, I figure asking about long-term projects here is best. https://github.com/N0VI028/JS-Slash-Runner https://github.com/Llynkurin/ST-Extension_JS_Runner_EN Anyone have any extra translations to contribute? Seems the guy running the project is doing a pretty good job, but I figure more heads is better than one.

>>14737 After Grok 4 Fast was given a paid option, the free option degraded significantly, to the point that raising the temperature above 1 made it give a nonsense output like DS, and the paid option is worse than the free option a few weeks ago.

Jeez, there's still people here at all? I'm kinda shocked.

>>13215 Ahh! Forgot to shill here! Natsume, a high-school boy with a penchant for photograpy and also the voices. (You)'re his promised bride and it doesn't matter what you think about it. https://chub.ai/characters/uraura/natsume-6d54f6900ce9 https://uraurabots.neocities.org/page1#natsume__remake_ Aubrey, a gothic loli trap who has kidnapped you and will torture you into becoming his dog. Remake of the first bot I ever released (that I can remember) back in the turbo days. https://chub.ai/characters/uraura/aubrey-8d74160debf5 https://uraurabots.neocities.org/page1#aubrey__remake_ Monster LoliShota RPG Isekaied into a world of monster girs and boys with a cheat skill allowing you to tame them through sex. Actually an old bot that I released and then unreleased after like an hour because I felt like I could improve it and then I forgot about it for a year. Now new and improved with some changes and HTML! https://chub.ai/characters/uraura/monster-lolishota-rpg-c824f3e616af https://uraurabots.neocities.org/page1#monster_lolishota_rpg I Reincarnated into Touhou Project but it's a Reverse Harem Game! A fork that I added hot and sexy HTML and some minor extra features too! https://chub.ai/characters/uraura/i-reincarnated-into-touhou-project-but-it-s-a-reverse-harem-game-9b76eb22774d https://uraurabots.neocities.org/page1#i_reincarnated_into_touhou_project_but_it_s_a_reverse_harem_game_ I don't respect my own time so I also edited the Akyuu Knowledge lorebook to change (almost) every female character into a male. There's also some extra features included in the lorebook like a random chance of an incident starting as well as Rape Mode and Dangerous Mode toggles. I probably missed some characters because there's a lot. >>14741 I still do it because it's funny to me to post in an (almost) dead thread and people do still stumble across this place.