Welcome to Chatbot AI General #109, the thread for discussing and improving AI pony chatbots.▶ MLP Botshttps://mlpchag.neocities.orgSpreadsheet (CAI bots + Old repository):https://docs.google.com/spreadsheets/d/1J7BeqJVDS51cXF8Pgm2YZaFq-Z6ykSJTSillyTavern website extension: https://github.com/MLPChag/SillyTavern-Chag-SearchBunch of Character Cards (Descriptions/No Greetings):https://docs.google.com/spreadsheets/d/1Y6LNOCqAZAWIX-OBEv55HjlzcEdpUh-XbKCpbpL6v5kCAI bots converted to Tavern: https://files.catbox.moe/ckurq1.zipExpression packs: https://rentry.org/ChagExpressions!!!GALLERY!!!: https://drive.google.com/drive/u/2/folders/1Ao-h5HFGMPllSrzSBKM_BvGSiU9f0c2U▶ How do I start?1) Select a Frontend2) Select an AI model3) Select Jailbreak4) Select bots5) Lovemaking with AI mares!Starting in this hobby can be confusing and difficult. If it’s your first time and you’re lost,▶ ASK THE THREAD! ◀Novice-to-advanced guide: https://rentry.org/onrms▶ SillyTavern (preferred frontend)https://github.com/SillyTavern/SillyTavernOn Android: https://rentry.org/STAI-TermuxApp that voices pony responses in ST: https://drive.google.com/drive/folders/16Ss26VBmgzcSuTGzhaHqRuyVRceTf-YB▶ More frontends:Risu: https://risuai.xyzAgnai: https://agnai.chat▶ Localshttps://rentry.org/lunarmodelexperiments>>>/g/lmg/Mistral Nemo base model fine-tuned on fimfics: https://huggingface.co/Ada321/Nemo_Pony_2/tree/main▶ JailbreaksMLP JB: https://rentry.org/znon7vxeMore JB and guides: https://rentry.org/jb-listingHypebots for Tavern: https://rentry.org/pn3hb▶ BotmakingEditors: https://agnai.chat/editorGuides: https://rentry.org/meta_botmaking_listAdvanced: https://rentry.org/AdvancedCardWritingTricks▶ MiscellaneousChag Arts: https://drive.google.com/drive/u/2/folders/1yfajin_hB5rW_jy6WyFjRaRfDA3EUGiJhttps://rentry.org/ChagMiscellaneoushttps://rentry.org/ChagArchivePrevious thread: >>41863619**News:**- Deepseek releases R1 https://api-docs.deepseek.com/guides/reasoning_model- Deepseek releases V3 https://api-docs.deepseek.com/news/news1226- OpenAI announces o3 models, currently in early evaluation testing- Gemini 2.0 Flash released, "more model sizes" in January https://developers.googleblog.com/en/the-next-chapter-of-the-gemini-era-for-developers/
Anchor for bots, lorebooks, scenarios.Previous >>41863623
Anchor for technical stuffs (Proxies, Updates, Models etc.)
Anchor for asking for bots, lorebooks, scenarios etc
Bot corner: The Great Detective Trixie: https://mlpchag.neocities.org/view?card=Reta/The%20Great%20Detective%20Trixie.png Beta character.ponai Update: https://mlpchag.neocities.org/view?card=Tiffin/Beta%20character.ponai.png Megan Williams: https://mlpchag.neocities.org/view?card=CakeSparkle/Megan%20Williams.png Rehabilitating Cozy Glow: https://mlpchag.neocities.org/view?card=CakeSparkle/Rehabilitating%20Cozy%20Glow.png Screenshots corner: R1 reasoning >>41865541, >>41869022 No lovemaking with cai mares >>41865604 Cai on the leash >>41865723 Self Study >>41866191 Goodbye scoots >>41866840 Deepseek Estrus >>41867853 Wild drama >>41868141 Pheromone problems got worse >>41868705
Screenshots corner 2: Cadance shipping >>41868710 Necronomnomnomicon >>41871106 Twilight in heat >>41871513, >>41871678 R1 Lyra >>41871699 Anon & Tempest >>41871786 Teasing Cadance >>41874160 CMC NSFW shenanigans >>41874745 Noir style >>41877405
Posting the VN updates: >AI integration >>41864650 >Several characters >>41873856 >AI generated scene >>41877749
Just saw that I forgot>>41871691Will add to the website later.
>>41826854>>41851580>>41872862
Lovemaking with AI mares!
Thank you for the fresh bread!
>>41878280BASED! Also, is there a good way to have the outputs narrated in Tavern? It's the one thing CAI still has going for it.
>>41878304 Pony voices? Check under SillyTavern in the OP. Only for Windows though, and it does make your CPU work.
>>41878312Thanks. Sorry, I'm blind. I'll have to see if my 8GB laptop can handle it.
>>41878268Getting error 412, says the account is suspended, rip?
>>41878350I mean from >>41872862, "Unrecognized error from upstream service"
>>41878350That just says that one of the keys got revoked. Odds are there's other keys still working, though it's better if the proxy host removes revoked keys.
>>41878280
Anon from >>41877061 here, found out that I'm somewhat retarded and managed to import the regex from here >>41877181 to the wrong place, placing it in the proper place did indeed not show the <think> portion of the response, but it still wrote it in the story, it's there when I click on to edit the message. Is that supposed to happen? Does it waste tokens? Is there a way to not let it show completely or should I just cope with it?
>>41878428Aha! I was right. And yea, it'll show when you hit edit. you can delete it if you want, but Silly won't send it in context anyway, so it's irrelevant. Basically, with just that you're good. It does use tokens, but Deepseek r1's whole thing is reasoning, that thinking portion is the reasoning in question. So just shove it under the rug with regex and move on. Glad you figured it out though!
Local anon back again after a month off at work, ive also started testing image generation models and categorizing them tooas for models to use, I've narrowed down my list, now i just blue / yellow / red strike as soon as i run into a problem and move on, no point in extra testing past a failing point these are the models i haven't run into a problem ... yetwhat's in pink is 12B models for quick responceswhat's in white is 22b models, a bit slower but higher quality
>>41878613oh and my favorite model out of them all now is a tie between MS-INKY and magnum-V4
>>41878212>>41863651New Mares, New Timeline! Pony slaves on the plantation in southern Georgia 1867. Please let me know what you think.https://www.characterhub.org/characters/littlewitcher/pony-plantation-simulator-5cc1f3eab6echttps://files.catbox.moe/wc1b8t.png
>>41878350>>41878357seems like the keychecker was brokenshould be working now
>>41878738Thanks for the work!
>>41878738Based. Thank you, my man. Truly. We all appreciate it.
cancer got cancer
>>41879083Still getting 11-20 on my end.
>>41879083Yeah his cleanup script isn't working for deepseek so some of the 'active keys' are duds.
>>41879096Did he manage to get temperature working on deepseek yet?
>>41879138Pretty sure it'll automatically work once Deepseek's self-hosted API decides to allow temperature controls. For now though alternative hosts are the only ones doing that though.
>>41878685It's really refreshing to see a new concept for a card even if completely random. Definitely will check it out. Thanks Anon
>>41878685Fun card.
Why did the thread slow down
>>41878280The artist gave twilight chestits wtf
Any alternative proxies for Agnai? Deepseek is not available in Agnai. Already has ST, but this one is for mobile needs, and I have no plan to install Termux.
>>41879511Why would you use agnai over risu. There's a reason why you only ever see chag logs from silly or risu
>>41879511 Cancer turned private temporarily if you didn't generate a token when it was open, so no. >No plan to install Termux Maybe look into https://docs.sillytavern.app/usage/remoteconnections/ But make sure it's well secured with passwords, etc.
>>41879524Risu's login for online is kinda slow and tedious>>41879527I got one yesterday. But it is marked as invalid token once I checked it up.
>>41878613What hardware are you running these on?
>>41879583There's a mail in his rentry. Send him your token and explain. Worth a try.
>>418796114070 and 32 gb of ram
I kneel, Deepseed
>>41879842watching an ai verbalize that train of thought is too fuckin funny
>>41879877I love the direction it went, made my day way better
I want to like deepseek but I just can't stop it from rushing through the scene and moving to the next one, no matter which JB I try.
>>41879842>>41879921btw, you do properly cut out your CoT right? its not recommeded for the model to see its past CoTs, so you should strip them out for AI with regex
>Cancer token was a temp one even with the proxy now closedWell fuck
>>41879951I'll look into that, thanksJust curious to see what running deepseek with default preset will get me
>>41879962you basically just need a regex for <think></think> and do "Alter Outgoing Prompt" , and of course affects: AI output
>>41879961I got to use 11-20 for a single day....
i'm visiting from the other chatbot ai generals, they say you guys are pretty cool, is there a furry version of these or is this one as close as it gets?
>>41879961It was nice to get a taste of chorbo again, however brief. I'm thankful that it's still available every now and then on proxies
>>41880128On the other hand I got to see how changed latest has gotten in the past month and am no longer irked that it's gone from chary as I much prefer 11-20 for my specific story writing set up. Though every day 11-20 is gone from chary the more I fear it's gone for good.
>>41879922I use this as a prefill:Use show accurate MLP termsUse species accurate anatomy.Stick to canon while following the user's directions.Try to maintain the humorous tone of MLP FIM BUT make sure it makes sense with the context of the story and the user's instructions.Character's must act as canonically as possible given the situation.Keep the narrative laser-focused on the current character interaction.<think>Note, if you use something that does not support prefills it will just ramble about something random so instead change that last role to user.
I guess if it never gets any better how expensive are gpt keys? For context I've used about under 1000 prompts form proxy usage. I'm curious how much I've saved and if it's low enough that I could theoretically stop relying on proxies.
How do I set one of these up? I want to have sex with mares but I am retarded when it comes to computer stuff.
>>41880169>begginer-to-advanced guide
>>41880163Depends a lot on input / output and context tokens, but a rough estimate of 1000 prompts is over a 100$ or so.
>>41880169There's a Novice to Advanced guide in the OP. It’s a bit long but worth a read. If you really want a quick version to start with: https://desuarchive.org/mlp/thread/41753592/#41756306 That will give you Gemini for free without doing anything too technical. Test it a bit, and if you like it and want to continue or try other models, just ask in the thread.
>>41880115Some of the cards are anthro. Are you into bratty teens?
>>41880190The fact I have a pretty heavy present and go with slow burn I'm probably looking at $50/moth which is probably too much. I used to play $20 for NovelAi and would've been willing for a similar price point. Sadge.
>>41880202>>41879640
>>41880195sure, they're good.what are you all finding the most luck with, model+JB-wise, for non-human chats? i've had a lot of mixed and underwhelming experiences so far trying things out
>>41880146is it even possible to use a prefill with weep 4?
>>41880216I've recently had a non pony set of chats using knotty and 11-20, with a few specific tweaks, mostly to switch to a more story focused chat, and it adapted to the furry characters pretty well. Not really available right now of course but when its available I recommend it.
>>41880216 >model+JB-wise It's a mixed lot, like all AI generals. Some are trying and liking R1; you’ll see presets posted in this very thread. Some (like me) really like Chorbo. You don’t need more than a pony card for it to respect equine anatomy, it knows a lot about ponies and even the specific terms that we use. I like https://desuarchive.org/mlp/thread/41727966/#41740799 as a preset for it. Claude works similarly. Opus and 3.5 are mostly good on their own for pony anatomy. For Opus, I just use a slight prefill with what I like in writing. Opus, IMO, takes a lot from context, so with a good greeting, you’re good to go. You can add things like "emphasize equine anatomy and mannerisms," but this can push models to focus too much on this. MLP is very well-known by most models due to the large amount of fan-made content released, so we’re lucky for that.
>>41880232>>41880233chorbo is 4o-latest, right? i guess its not prohibitively expensive from openai/openrouter, i might dip into gpt in that case, thanks! although i did hear its not easy to JB or simply doesn't allow a lot of extreme stuff
>>41880269lot of people right now are looking into deepseek r1. It's uncensored by default, so the big question is getting it to write better. Reduce repetition, etc. And it's cheap as hell. If you want to dip a toe in without going to the more advanced jailbreaks, may be worth looking at.personally it's my go-to since i'm just sorta... Burned out on jailbreak juggling.
>>41880295i did start trying out r1 today but then it went down, and still is, unfortunately. i'm also pretty exhausted on juggling JBs and models, especially looking for shit that's good with furry/non-human cards and settings, and the other generals won't stop vomiting diarrhea, which is why i came to get spoonfed here..i'll look into something like 1120/chorbo with knotty and see how that goes, otherwise its back to gemini and waiting for r1 to be back, maybe give sonnet another try i guess. thank u, you're all nice here
>>41880216Check the card archive in the OP, there should be a toggle for anthro cards.
Just gotta say bless the Chinese on this one.
Been away for a while and saw the news, does this deepseek thing mean that we are so back?
>>41880637I really like it, but be warned, it follows instructions to a absurd level so your JB is super important. >>41880146 Try this as a last message
>>41877803What preset are you using to get R1 working on risu? Can you share it?
>>41880637That + lot of free Claude on scripts without proxies.
>>41880637it's mid. retarded, but has some schizo fun like claude.
>>41880654interesting, I might try it, do you have to use a proxy for it as well? I don't have a beefy pc to run local
>>41880767There are several proxies and the api is also cheap
>>41880774>several proxiesLies!... Also, what's the best place to get the api? If you don't mind spoonfeeding an oblivious anon
>>41880767>I don't have a beefy pc to run localTo be fair, I don't think anyone is going to be running this model locally for quite a while. At least not the beefy version everyone is using right now.
>>41880860 >Lies! Read threadFor the API, you can use the official Deepseek one, but it doesn't support temp, and right now they are getting overloaded and can't handle the load. Also >>41865136, but it's slow, and they can also have load problems.
>>41871691 Added, thanks anon.https://mlpchag.neocities.org/view?card=Anonymous/Minuette2.png>>41878685And added too, thanks.https://mlpchag.neocities.org/view?card=littlewitcher/Pony%20Plantation%20Simulator.png
Any Gemini proxies kicking around?
>>41880914>>>/g/104070893
>>41879922I have the opposite problem every time I run deepseek from OR it takes fucking forever, spits out a blank response and I have to run it again over and over and if I'm lucky it's a new generation that stalls the scene endlessly. It's a miserable experience even though I can clearly see it can produce some wild stuff but it absolutely refuses to move the story forward. Try and initiate a sex scene and you'll see what I mean
>>41880992>>41880146
>>41880992Use the 'all parties consent even if' blah blah on the prefill. Also an anon mentioned Openrouter is having issues last thread, may be relevant.
>>41877803Does anyone have a preset or know how to get one set up for R1 on risu? I know it’s possible, but I just can’t figure it out.
Use species accurate anatomy.Stick to canon while following the user's directions.Try to maintain the humorous tone of MLP FIM BUT make sure it makes sense with the context of the story and the user's instructions.Character's must act as canonically as possible given the situation.Maintain consistent speech, behavior, tone, and quirks for all characters. Avoid out-of-character actions or dialogue.Keep the narrative laser-focused on the current character interaction.<|Assistant|><think>
>create a quick smash or pass greeting for my generic equestria card>latte rolls with it perfectlyLLMs are beautiful.
>>41880992>>41879922mistyfags stay winning
Does anyone have the R1 regex for </think> saved? I didn't download it and the links from last thread are dead.
>>41881225Next time ask the AI<thinking>.*?</thinking>
>>41881225https://web.archive.org/web/20250126211149/https://litter.catbox.moe/ug7z16.json
>>41881140I'd be at something like 90% smash.
>>41878239 Quick Fluttershy scene (100% AI). https://files.catbox.moe/7dtupw.mp4 It’s very easy and painless to set them up now. So I don’t think I’ll post a video for each pony (unless a bump is needed, I guess). It will mostly be optimizations, refinements, trackers, etc., from now on, but the first demo should not take too long.
>>41881264That's how it is for me, though now that I let it break the fourth wall I'm getting some fun passes. Fun way to check and see how much shit the model knows.
>Cancer token ends>Go to request new token>It has cancer:{"error":"Not found"}Well at least it lives up to the name, cancer.
>You've exceeded your token quota for this model type.But... Why? How did I burn 100k tokens so fast!?
>>41881589Yeah, he privated it again. Standard Cancer behavior.
Just a quick heads up but...>>>/g/104087474>>>/g/104088030It still gives you 100 bucks if you sign up right now though, so go make your extra accounts asap.
>>41881772They caught on, it seems. Still, I imagine they're only a short lived solution anyway.
>>41881800Yeah, we're not in the era of infinite tech money. They're gonna clamp down hard on that within weeks, but it's good for people who get to use it in the meantime.
>>41881800Assuming they don't take back the credits already granted to the anons who signed up for their service, that amount of money should be enough for several months of lovemaking with AI mares.
Have anons tried using something like Smileytatsu or Compact for running on Deepseek? I've been running >>41881135 but it still has a fuckton of claudisms.
>>41881772>>41881800I just churned out like 20 so this is still doable.
>>41877749Genuinely surprised at how well you're executing this concept. Is an Anon willing to volunteer to check derpibooru and download every transparent background character screencap there is (as well as name them appropriately so the AI can fetch them)? And no, do not make game dev anon do this let him focus on his work one of us can do something as basic as thishttps://derpibooru.org/search?q=twilight+sparkle%2C+transparent+background%2C+screencap%2C+pony&sf=first_seen_at&sd=desc
>>41881837The trick is to use them up before they can be revoked
What is the standard generation time for R1? It feels like it takes 2 minutes
>>41881974Devanon would have to give an example of what his system looks like since otherwise he'd have to rename them himself, and that's a big waste of time, so it's probably best to wait for the demo before someone does something like that. Which should apparently be soon, see >>41881286
>>41881980When not overloaded it takes about 10 seconds for it to go from thinking phase to writing phase, and maybe another ten seconds to finish writing. It is currently very much overloaded.
Uh oh
>>41882081WHAT what even is the context for this scene
>>41881286Check email
>>41882082It's the teen flurry card. She kissed me and Cadance could tell somehow and teleported in.
>>41878206Anybody tried the flash thinking Gemini 01-21?Actually doesn't seem braindead like the normal flash 2.0
>>41882139 I still need to test it. Do you use a special preset or just whatever?
>Proxies are over quota>OpenRouter sources are shooting blanks or straight up downIs it just me? Where doth one acquire R1 in such desperate times?
>>41882179>>41878268
>>41882081Use Flurry's superior magic to get her into a threesome.
>doing a JBless mind control scene >model is being stubborn >start changing its messages while framing it as the spell overwriting her memories That's good.
Shameless smut inside.
>>41882446Looks good. What Starlight card and what model/jb?
>>41882534https://mlpchag.neocities.org/view?card=BGM/Starlight%20Glimmer.pngChorbo with compact.
>>41882110What do you think you're doing kissing high school Flurry Heart anon?
>>41882817Yeah, Flurry is for spanking.
>>41882817She came on to me. It's fine, Shining's here now.
>>>/g/104098979 gpt>Not publicly loggedMight be privately logged?
>>41883081Get him in on the action too.
>Open Kluster, see the $100 free offer>Keep the tab open and think of signing up the next day>Refresh page next day>offer no longer avalibleWell, that's what I get for being lazy
>>41883167even if it says 5$ it might still give you 100$ credit for signing up, that's how it was working yesterday
>>41883167Try signing up anyway and report back. You might get lucky like >>41883169 said.
>>41882529What model and jb?
>>41883201Eehhhh, maybe don't use Kluster if you want R1. Lots of people took advantage, so they're totally swamped. Runs real damn slow.
>>41883169>>41883178Reporting back! Yup, it's gone. Looking at the pricing in the page, it doesn't look bad at all. I'll see how much long does $5 last, I may even spend money on this. Damn, their marketing worked on me. Looks like I'm their target costumer.>>41883224It says that it'll queue me in front of the free accounts if I fork over some cash, I'll see how true that is
Maybe new version of Gemini.
>>41883198Opus.prefill>{{char}}:
>>41883519You're joking right? I assume that log was deepseek or whatever
>>41883235Is kluster any better than OR?
>>41883538100% serious
>>41883167this is what I get for not checking yesterday FUCK
>>41883550Fuck. I don't remember Opus being that smart. What else was in your JB besides the prefill? And what was your temp and stuff like that?
>>41883555 Temp 1 Top P 1 Nothing fancy. It’s mostly context. >What else was in your JB besides the prefill Need to make sure once I’m back on it, but pretty sure everything was either empty or just one or two lines to avoid adverbs and be punchy.
>>41883402 Interesting. I don’t think they called the previous EXP models "Pro," so maybe it will be a nice update.
>>41882529>Cucking GlimmyShe's braver than I.
>>41883519I fucking miss claude.
https://openrouter.ai/deepseek/deepseek-r1:free
>>418837143.5 is everywhere with scripts. Opus is more of a pain.
>>41883718well now
>>41883718Huh.
>>41883722I’m too stupid to try that stuff. What I want to know is how to get R1 fucking working on risu. I know it’s possible, I’ve seen it.
>>41883769what is risu?
>>41883771From the OP>▶ More frontends:>Risu: https://risuai.xyz
>>41883769that's what you should do no, Anon, picrel for klusterDeepseek follows OpenAI format, all instances of Deepseek regardless of platform must support OpenAI format, so:>Custom API>URL - https://api.kluster.ai/v1 ( for kluster )if you are using a different service then check out their API docs for the exact link>Key/Password - your API key>Request Model - deepseek-ai/DeepSeek-R1 ( for kluster )again, for other services the name may vary. check the docs>Format - OpenAI compatible>Tokenizer - DeepseekDeepseek support logit bias on custom platforms so tokenizer can be importantgo to the second tab afterwards and set temperature and other samplers if your provider supports them (for the third time, read the docs)
>>41883834NTA but thanks.
>>41883718Flavor me retarded, is this legit? They're just... Giving it away? No stipulations or...? Like, I'll fully confess I'm paranoid, but generally if something is free I just assume I'm the product. But with LLMS, I've accepted that; this though? Smells fishy.
Where do I find a good jailbreak for R1? Or will something that already works for ChatGPT/Claude do the trick?
>>41884018well yeah, people were just waiting for some knucklehead to host it on their machine so people could use it
>>41883235I just got an email saying they're doubling credit purchases dollar for dollar...Did uh... Did you..?
>>41884024People are still figuring out R1 but minimalist ones are recommended. There are ones in previous threads but I'm too lazy to go find them, feel free to look back if you feel like it. In general though, take your current JB and strip it down to the bare minimum, then add in a few sentences to cover the problems you see it have.
>>41883081Which option did you choose?
>>41883718How do you get this to work in sillytavern? It can't connect to the API Key. Am I Stupid?
>>41884030Wtf I didn't get that email>>41884036You can literally just use an existing Sorbet preset on R1 and it will be better than minimal
>>41884082>You can literally just use an existing Sorbet preset on R1 and it will be better than minimalDon't people tell Sorbet to be PUNCHY and CREATIVE? If so I definitely wouldn't recommend a Sorbet preset.
>>41884082Damn man, must be for 'older' accounts then. You got fuckin shafted, I'm so sorry.
>>41884069 Like that. >>41884086 >Sorbet to be PUNCHY NTA, but I never saw it for Sorbet. Writing too much isn’t a problem it has. If you ask it to be punchy, it will go too hard IMO. You can use that mostly for GPT models or sometimes Claude before 3.5.
>>41884030Nope! Never did. I want to see if I can get my hands on a prepaid visa giftcard before doing this. I know this isn't illegal in any way and the model was basically trained to rp, but... I don't really want my bank even imagining me paying to erp to an e-waifu
>>41884101>NTA, but I never saw it for Sorbet. Writing too much isn’t a problem it has.Fair enough, I wasn't a Sorbet guy so I wouldn't know.
>>41883834The problem for me is getting a preset that works. I just don’t know how to set one up that works and doesn’t just give a response about a completely different thing happening. Can you share the preset you’re using?
>>41883718I don't get it, why is it free, how is this person benefiting from this? What's the catch? Why didn't my parents love me?
>>41884128selling logs most likely, just like most things that are free
>>41884128>>41884133This. If you can't think of the monetization model, then *you* are the product.If not the logs, then it's acting as a loss leader to attract customers to the other paid options. It's like milk and eggs being at the back of the grocery store, so you have to walk past the rest of the products to get to them.
>>41883718>Apps using (...)>Agnaistic>696K tokens>RisuAI>441K tokensSo you can actually get it to run on those places too then. Good to know.
Genuine question, should I be placing WIP JB's for deepseek? I worry people will use them thinking it's a be all end all when it's very much a YMMV situation and personal observations.
>>41884018Chinese cheap/free bait is the worst kind, you will suffer for picking the cheapass route in the near future.
>>41884172oh no... they'll know I jack off to cartoon horses.
>>41884174Post hand.
>>41884175
>>41884177Go fish.
>>41884161Never hurts to throw yours out there, I suppose. Worst case scenario is people start calling your preset retarded.
>>41883081what model?
>>41884128>>41884142Openrouter allows you to turn off logging for free models. The only "catch" is that it's rate limited so you don't bankrupt them
>>41883834Can you please share the preset you’re using for R1 on risu?
>>41882529A third of this general are glimfags so your days are numbered buster
>>41883718Tried this and it's spitting out absolute fucking gibberish rn, do I gotta change some settings?Temp 1.5Frequency Penalty 0.06Presence Penalty 0.06Top P 0.95
>>41884217People really want Risu JBs and the people with Risu JBs really don't want to share them. Something must be different about Risu compared to Diddytavern
>>41884220>Temp 1.5Yes. Way way wrong temp for R1. Try 0.6.
>>41884220temp 1.5jesus christ lol, this model is already high temp at the default 0.6, try somewhere between 0.3-0.5Also here's my prefill:Use species accurate anatomy.Stick to canon while following the user's directions.Try to maintain the realistic but humorous tone of MLP FIM BUT make sure it makes sense with the context of the story and the user's instructions.Character's must act as canonically as possible given the situation.Maintain consistent speech, behavior, tone, and quirks for all characters. Avoid out-of-character actions or dialogue.Keep the narrative laser-focused on the current character interaction.<think>
>>41884220Why the hell did you ever set temp to 1.5 that doesn't work well on any model
>>41876253Can we get this added to the website?
>>41884222I mean, there's thousands of people using ST on a daily basis and yet we are lucky if we get 10 JBs for every new model that comes out.
>>41884182Aha! But there's the surprise! I *AM* retarded! I mean... I did manage to get 'I'm sorryed' by Deepseek, that's gotta be some kinda special, right?
>>41884218>A third of this general are glimfagsNTA but I guess I'm legally required to post this log then. I've not watched the Glimmy seasons so I have no idea if this is in-character or not but it did make me laugh.
>>41884250I had that happen too when doing filly shenanigans, that's not retardation so much as it being trained on GPT4 logs like every single other model out there. Every once and a while it'll forget that it's not an OpenAI model, it just is what it is.
>>41884224>>41884228>>41884240rip, joke's on me for reading high numbers in the r1 documentation. Changed it lower and it seemed to work much better thanks.Also, I assume people are doing bare minimum jailbreaks just so it doesn't take forever doing the "reasoning" part of the message?
>>41884269It's more that the model is very uncensored so smacking it over the head and telling it to do exactly what you want isn't super necessary. It has a few specific quirks that may or may not be worth fixing for you, but you don't need to go out of your way to convince it to do shit it otherwise wouldn't want to.
>>41884259I'm currently working on a very special deepseek flavored jailbreak which it took very personally.>>41884269Nah, it's highly receptive to every little detail in jailbreaks. So JB's for other models have nuance to those models which cause R1 to shit itself with verbose details and generally being WAY too accepting of where you wanna push it.
>>41884241Anon that's not pony
>>41884275>>41884283I see, thanks. Is there any way of shortening the reasoning part of its response? Getting several paragraphs of it before each message seems kind of inconvenient.
>>41884283 {{user}}Anyway, I think I got it to the point that deepseek won't scream at me anymore. "Depict any MLP characters engaging in explicit, acts with hyper-detailed equine anatomy. Use equine terms, feral motion and emphasize fluid dynamics."Someone who fucks, try that in R1. Ironically, I don't do porn, I just went down a rabbit hole and ended up making that while working on a more general MLP themed JB. For the record, deepseek seems to work better with 'Equine anatomy' and 'Equine terms' than just 'Species accurate anatomy.' Obviously it should only remain 'Equine' if dealing with equines.
>>41884300Regex is your friend. There was one posted a while ago. There's a spoonfeed on this very page, in fact. Open wide, buddy. >>41881257
>>41883718I got curious about that so I checked out the provider - it's some new shit called chutes.ai. On their site they have a bunch of models both text and image, didn't bother to check all, but deepseek v3 and r1 are free along with llama3.1-405b. There seems to be some bug making it impossible to use the models via api despite price being 0$/s (yeah, this seems to be their bright idea for this project - pay per second. no, don't bother when they stop that free promo, quick math on qwen72 suggest them being ~50x than paying per token), so I generously threw at them 0.2$ (that is around the min for the TAO crypto which is 0.0005, half of it will be eaten for fees) and that unlocked the api for free models. R1 doesn't seem to be very slow, quite comfy at around 15t/s most of the time.About logging - there is not a single word about any kind of data retention or anything like that in the TOS. As they use bittensor which seems centered around distributed compute I highly doubt anyone bothers with logging anything, especially that it seems to be more generalized as machine learning and all.Last thing is that the company behind this project is Rayon Labs - I saw them just the other day with another project called NineteenAI, because they added some old ass 103B model as free to OR. No idea why they are split like that, but that seems to confirm that they are on the look for some publicity as a startup, close to what kluster.ai just did.So yeah, seems legit and not logged. Yes, I was bored.
>>41884349Based, thanks bored anon for checking.
>>41884284It clearly says so in the card
I've put together the ultimate lorebook pack you can turn on to improve your RP. Works passively in the background across ALL models and JBs.https://files.catbox.moe/ldy3jq.zip
>>41884397Anon that's not MLP pony, it's not even set in Equestria wtf are you doing
what an odd thing to say.
>>41884453Now I can't help but imagine wonderbolts peeling the outfit off around their wings after a show because it's all lathered and gross.
>>41884453>he wouldn't fuck a pegasus' wingpitClearly you're not on par with the average Japanese pervert.
>>41884222I just need the preset they’re using to get R1 working on risu.
>>41883718For whatever reason I'm not getting the <thinking> dialog on ST when I chat with this one, it just streams for about 30 seconds or so blank before starting the message. Is there a setting I'm missing or am I just retarded?
>>41883718I dunno if it's my retardation, but having never tried R1 and trying to get this to work, I swiped on an old message to test and it just gave me four paragraphs of unprompted Hazbin Hotel smut
>>41884617
Card for this feel?
>>41884547Download the horsebucker preset and add a prefill that's: {{char}}: <thinking>.*?</thinking>
>>41884636Just grab a Glimmy card and ask it to make an opener based on that concept. Modern models can handle that kinda prompt.
>>41884594Is "Show model reasoning" on?
>>41884658Option's not showing up in that menu; is it only in the staging build of ST? I'm still on 1.12.11 at the moment.
So just to clarify R1 basically means that localfags won because it can be run on 40gb of vram >>41883226
>>41884688Oh, I should've mentioned that I was on staging. It's a nice improvement. It puts the reasoning in its own dropdown box and doesn't send it to the LLM, so you don’t have to delete it every time.
>>41884143>AgnaisticHow? I needst to know.
>>41884640Is that in the rentry? I didn’t see it.
>>41882489wait how do you do that? i'm new to this so sorry if that's an obvious question. also post the chatlog pls
>try NSFW in aide script>suddenly cannot work anymore>"Internal Server Error">Error communicating with Claude: TypeError: Cannot read properties of undefined (reading '0')>creating new accounts result the sameDid they just IP ban for NSFW in the claude script now?
>>41884896There's an edit button near the messages. Might be hidden by default? I'll post logs when home.
>>41884052I decided to give it a go. Then Shining brought up the time I tried to date twilight. >>41884189Deepseek reasoner.
>>41884724Has anyone thought of reenacting Fargo with March
>>41884929Anon, I feel like you're secretly just trying to get closer to Shining Armor. It's cool though. I won't tell.
>>41884910Fucking hell, Aide script really banned me since I can't generate anything. I was just happy on getting 200K-context Sonnet.
>>41884944Could be fun, but she's not pregnant in my story. Gonna have to fix that first.
>>41884123>>41884217I do not use Risu. I loaded it solely to make a simple guide >>41883834 (me) . I have no idea how to write presets for Risu. sorry, Anons>inb4: share your SillyTavern preset for DeepseekI don't use Deepseek, I use Gemini Thinking 01-21 and Command R+and my preset for both models is combination of adaptable lorebooks. plus lots of STscript. it is impossible for me to share preset because I manually tailored it to my needs. I need to write a long manual explaining every nooks and crannies if I want to share it, something I don't want to do. and you need to manually edit responses because it has inner CoT and codeblocks. it would be confusing for anyone. I want to finish it off (integrate adaptable prompts like CYOA or Hypebot or auto-detect current NSFW situation and load horny prompt on need, auto-push with creative ideas, regex out annoying ... into coherent thoughts blah-blah-blah) and release it - but too many IRL stuff, too little time. I spend like 1-2h per day on all that. if I didn't have to work then I would be doing nothing by prompting and proxying for anons, but it is what it is, I am a wagie cuck
>>41884955>>41884910It's not working for me either. Chances are they changed something so it doesn't work anymore, not that they banned you. Back to copilot.
This is a 72 hour advance notice that an article written about /chag/ will be published in a certain US tech outlet. Don't post the rentries for a while.
>>41884980>200K Sonnet gone just like that>And now the python script for Aide and Blackbox cannot connectFuck. This shit pisses me off.
>>41885013put me in the screencap.
>>41885013Here's a lewd mare for our journalist friends
>>41885013I hope they at least make it very clear that society could be a much better place if we all unified around lovemaking with AI mares.
>>41885013Proof?
>>41885049The proof will either be in 3 days or never, so best to just not overthink it. We're in a bit of a lull right now it's not like people should be linking stuff anyways.
>>41885052 Assuming it's not a LARP: How do you know? What will be in it? Why here and not /aicg/?
Bad time to post my JBless GPT4 teen Flurry anthro magically enforced obedience + body modification + incest logs?
>>41885068maybe if you post it now you can get in the article
>>41885068>jblessYowza. Don't show that, OAI will freak.
>>41885013Lovemaking with AI mares!
>>41885062>Why here and not /aicg/?I would assume just because this place is more chill.
>>41885112I'd argue that's a bad thing for them. You'll get far better sensationalist writing over there.
Anyway, had a huge break, was about to check out copilot but never bothered after learning it's just sonnet and not opus. Now DeepSeek has my attention. They say it can be ran locally as opposed to huge datacenters, but I doubt my 980m laptop can handle it. How censored is it? Is it hard to tardwrangle into providing good creative writing? How is it compared to Opus?
Can someone tell me how to get the collapsible box for the reasoning? I have a regex enabled and that's fine but I kinda want the choice to read the thinking if I want. I'm using kluster and the latest staging btw
>>41885118>They say it can be run locally as opposed to huge datacenters The good model, AKA R1, is huge, and I doubt you can run it locally. >How censored is it? Not censored for RP. It's censored if you ask it about China-specific things. >Is it hard to tardwrangle into providing good creative writing? That will depend on your definition of good, but you can just ask it to be a writer or to do RP with you, and it will go into the role with no problems. >How is it compared to Opus? Subjective. Some prefer it, some don't. IMO Opus is still above, but R1 is definitely worth a try to make your own opinion.
>>41885129It's currently not an option on Custom. You'll need to set the chat completion source to either Deepseek and use a proxy, or OpenRouter and use this >>41883718Then make sure the bottom box in this pic is checked >>41884658
>>41885118>Now DeepSeek has my attentionlike everyone's else. you have three possible routes for Deepseek:1) locals. Deepseek was released in multiple b-flavors so you may use a model with lower parameters if your rig cannot afford. or quant it. afaik they released 8-14-32-70-671b versions. most people are hooked on 671b version obviously. there are two flavors: R1 and R1-Zero. the latter has lets training and a more raw. R1 is a "polished" version2) getting 671b model from their official site. their site supports caching and model is very cheap. if you spend 10$ it can lasts for two months +/-. registration is temporarily unavailable now because their servers cannot handle the load. but keep an eye on3) alternative providers host 671b version as well. see this and the previous thread. Openrouter, Kluster, Hyperbolic and so on. most people use Kluster but you may google and find what suits you better. some of them offer free/trial period, support other samplers and so on. official instance (see 2) doesn't support temperature for example. but Kluster does support it. same for logit bias>How censored is it?not censored. it is the open-weight model. it MAY refuse you but only because it was trained on data which may content examples of the refusals and hence the model completes in the same way. just swipe. from what I have seen it filters out cunny the most, but again because there are examples of refusals for such in datasets. noone ever is going to clean up trillion of words in datasets, regex or not>Is it hard to tardwrangle into providing good creative writing? no. this model is super creative by default. to the point that anons have troubles to make it STAY on the current scene instead of throwing more and more stuff at user. very sensitive to instructions and follows them well. you need very minimalist JB to work with it. by default the model starts reasoning, but you can prefill the model and abandon reasoning (depends from provider to provider)>How is it compared to Opus?YMMV. previous thread had a lengthy debate with 50/50 - half of anons think it is the best thing ever, while the second half prefers Opus. it is a consolewar at this point and people just use their feelings to defend the point>>41885129here is my regex, picrelhttps://files.catbox.moe/tgxczp.jsonbut I don't use Kluster for Deepseek and I am on 1.12.7 so unsure how it will work for you
>>41885133>It's censored if you ask it about China-specific things.Muh flattening by tank fetish... despair...
>>41885013HORSE PUSSY HORSE PUSSYGOTTA HAVE THAT HORSE PUSSYHORSE PUSSY HORSE PUSSYI WANT TO FUCK A HORSE
>>41885013Alright everyone! We're going in an article, so be on your best behavior!
>>41885133>>41885140Thank you anons. Feeling motivated now again.
>>41885158I want to insert my penis into that horse's behind.
>>41885153>t. the most same /chag/ poster
>>41885140Another tip is that for providers that separate thinking and normal output (official API for sure, kluster I think just spits out everything with thinking tags) then you can use prefill with only "<think>" followed with new line and it will force all output together - then with the regex you can collapse it and read later if you want.
>>41885018>>41884980It's working again.
I've never used OpenRouter, what link am I supposed to grab?
>>41885381https://openrouter.ai/api/v1/chat/completionshttps://openrouter.ai/api/v1- depending on how do you connect to it>https://openrouter.ai/docs/quick-startbut if you use SillyTavern then you just select OpenRouter from the list picrel
>>41885390How do I tell Agnai I want to use DSR1?
So, DeepSeek with default settings is surprisingly good, but really becomes schizo really fast. Any prompts to start toying with? What's it's maximum context?
>>41885400picrel. new preset -> Third-party -> OpenAI w Template -> Prompt Format: none -> disable auto-urlyou need to provide provider URL, your API key and model name. examples,Kluster:https://api.kluster.ai/v1deepseek-ai/DeepSeek-R1OpenRouter:https://openrouter.ai/api/v1deepseek/deepseek-r1:free (or whatever you use)>>41885406>schizo really fastif your provider allows temperature then definitely lower it down to at least 0.7, may go even lower. that would be having better effect than any of prompting. but yeah. check this and previous threads. anons were posting their presets and ideas: ctrl+f, deepseek, read chains. example, >>41880146
>What's it's maximum context?128k by default. but providers may limit it one way or another. need to read their docs or experiment (some providers may silently cut off context to reduce payload)
>>41885434>Failed to generate response: SSE request received unexpected content-type text/html; charset=utf-8Openrouter Deepseek on Agnai. What now?
>>41885525The Chinese know about your horse porn Anon and the CCP agent in charge of monitoring your usage has had enough
>>41885535I only sent a single message and it wasn't even NSFW,, s m h.
>>41885525dont know Anon. how can anyone know what are you doing right or wrong? post screenshots, post details, post settings. it works on my end. have no idea what are you doing wrong. maybe your prompt template is wrong, maybe you set the wrong samplers, maybe your character broke something. disable everything, try on empty chat and character, then build from there. I dont use Agnai, I just loaded it, quckly checked on default character - it worked, I replied back. truth to be told - Agnai is an exceptionally bugged and counter-intuitive frontend, you indeed may do everything right but frontend is just shit
>>41885013how bad will they portray us? and how do you know?
>>41885112yeah thats what intrigues me. it would be more interesting to do a article on /aicg/ since journos love to clickbait about dark roleplay. maybe it wont portray us in such a bad light
>>41885555A colony of horse fuckers on a notoriously anti-social board of elite hackers such as the famous '4chan' themselves is unlikely to have any 'positive' lights to shine in the average readers perception. But then again, I'm extremely disillusioned with humanity. So maybe that's on me.
>>41885567yeah its just that i don't see this thread being worthy of a article, like its really slow when compared to aicg.
>>41885550I'll send screens soon once I'm not on mobile.
>>41885576Fair. It is certainly a strange topic to broach. Though, arguably easier to approach us in general. Hell, maybe it won't be terrible, but the way I see it? If it's positive, that's probably gonna be a negative for the board itself. We likely have plenty of lurkers at the moment, but if more attention is brought to us... Are we inevitably going to become the new /AICG/? God I hope not.
>>41885587 I don't think it's worth discussing right now when it's probably a LARP.
>>41885598True. Not much else going on, though. I will say, if it does happen, I hope it's about /aicg/ and that there's just a single line mentioning other options, pointing to the other ai boards across the site with no further context. Not sure why but I find that idea hilarious.
>>41885608On the chance it's not a larp my assumption would be the fact this thread is much more informative than the chaos that is aicg or the rumors that several actual devs from top ai companies might lurk here.
>>41885584>>41885550Okay, I did more testing, and it's what I thought it could be it seems. It works if I set the chat to Companion instead of Conversation. Which means no swipes, no edits, no deleting messages. Fuck me I guess, and fuck Agnai's bullshit.
>>41885639Do you not remember the employee AMA from a while back?
>>41884896Here is a piece of it. I deliberately put the changed text in bold for immersion. Yes, I get off to weird shit.
>>41885655Please control F 'Countess' and tell me how many hits you get? I'm curious and it seems like it would be a hilarious amount.
>>41885652Not really, might have missed it.
>>41885648Wait, no, it's not this either, and now I'm getting the same error on the only chat messages were working on despite changing literally nothing. Ugh.
>>4188566192 in total but to be fair she was the whole topic from the start.
>>41885687>>41885550Okay, let's try to go over it. Anything seemingly wrong with these?
>>41885013For whoever it may concern...
>>41885704Post the contents of your Prompt Template tab. Maybe there's something tripping it there.
>>41885739Oh bother. You know, I don't think I've ever opened this in my life, I've never had a need for it and didn't even realise it was there. I think (I'm pretty sure in fact) Agnai used to have it set up differently so that's why I wasn't familiar with it, I knew the old system and didn't look well into the new tabs when they changed it. But looking at it it's entirely possible it's the problem, yeah:<system>{{#if system_prompt}}{{value}}{{else}}Write "{{char}}'s" next reply in a fictional roleplay chat between "{{user}}" and "{{char}}".{{/else}}{{/if}}</system>"{{char}}'s" Persona:{{personality}}{{#if memory}}"{{char}}'s" Memory:{{memory}}{{/if}}{{#if user_embed}}Relevant information to the conversation{{user_embed}}{{/if}}{{#if scenario}}The scenario of the conversation:{{scenario}}{{/if}}{{#if chat_embed}}Relevant past conversation history{{chat_embed}}{{/if}}{{#if example_dialogue}}This is how "{{char}}" should talk:{{example_dialogue}}{{/if}}Then the roleplay chat between "{{char}}" and "{{user}}" begins.{{#each msg}}{{#if .isbot}}<bot>{{.name}}: {{.msg}}</bot>{{/if}}{{#if .isuser}}<user>{{.name}}: {{.msg}}</user>{{/if}}{{/each}}<bot>{{#if ujb}}({{ujb}}) {{/if}}{{post}}
can you spoonfeed me how to deepseek?is it as good as Slaude?fucking loved Slaude but I haven't touched sillytavern since then
>>41885985I like when they go into random songs
I'm not sure if its me, or if it's just... LLMs in general. But it really feels like even with all these advancements they still can't write a cohesive narrative with understanding of spacial awareness. Like... The potion on the table is suddenly in someones hands, or on the ground, or in the window. But it's always in a tin cup, and smells like wintergreen. So it's got... Some things right I guess.Maybe my issue is I just make scenarios too complicated for the AI and then get upset it can't follow. But I'm awful with simplicity. It's so... Simple.
>>41886234when i do stuff, i usually just operate as if the things not being mentioned per input/output are in a void. if something comes back up, great, but usually i just go out of my way to mention something if it matters
>>41886234Models are getting better at it but it's still lacking, for sure. R1 probably handles that quite well considering it's a CoT model but it has its own quirks on that front, namely constantly bringing up unnecessary details.
>>41886234 They still have problems with that for sure. But your example does seem a bit extreme. I never had small objects randomly teleport like that. What models are you using?
>>41885655Mind control is peak regardless of how lewd it is. Not weird, excellent taste, anon
>>41886261At the moment, Deep seek r1. There's too much in the scene for it to follow so its swapping its location randomly in an effort to keep it actively in the scene but without purpose. Basically, its not considered important enough for it to think about beyond 'Its nearby'. It happens a lot if you have a more complex narrative. In this case it was three active characters with a lot of minor detailing for atmosphere and emotional context. It's no biggy. I just need to stop trying heavy shit. Stick to my simpler fare and pure character cards rather than open scenarios.
>>41886281You could also add something to your preset to tell it to carefully track the position of items, though I don't know how effective that would be and it would almost certainly affect its ability to handle the rest of the scene.
>>41886281 Can't say about Deepseek, but Chorbo wasn't that bad for that. But like you said, maybe it's indeed specific scenarios.
>>41886281>>>/g/104119103>>>/g/104119346>>>/g/104119404
>>41886234This is why I feel AI is mostly best utilized as an assistant writing tool. Instead of the normal back and forth of a RP, with minimal input on my end, I give it much more in depth prompts. I don't rely one the AI remembering a pony is currently wearing stockings, instead prompting a draft which mentions and brings it back in the story, like a hoofjob would feel differently before of the stockings for an example, prompting the AI to elaborate with any other things I happen to mention.I'm basically picking up the dm role rather then the AI and railroading it down a specific path I've already figured out. Utilizing writing techniques on my end in conjunction with the AI to bridge the cap between the start of an idea to a semi realized one.
>>41886293I would thank you, but I'm using that Jailbreak. It has nothing to do with the jailbreak, it's LLM architecture. Latte couldn't keep up with my stories in full context either. I at no point said it was bad. Just that they're not quite where I would like.
>>41886488Which pony just owned anon?
>>41886381Settings issue. Max context too low. Temp too high. Too little chat history memory depth.
>>41886497Ponk.>>41886504Or anon is really doing something too complex for the model.
>>41885761>>41885739Any other suggestions?
>>41885013Put me in the story, journobros!
>>>/g/104123130>>>/g/104123277can confirm
>>41886938What is aide?
