Welcome to Chatbot AI General #101, the thread for discussing and improving AI pony chatbots. ▶ MLP Bots https://mlpchag.neocities.org Spreadsheet (CAI bots + Old repository): https://docs.google.com/spreadsheets/d/1J7BeqJVDS51cXF8Pgm2YZaFq-Z6ykSJT CAI bots converted to Tavern: https://files.catbox.moe/ckurq1.zip Expression packs: https://rentry.org/ChagExpressions !!!GALLERY!!!: https://drive.google.com/drive/u/2/folders/1Ao-h5HFGMPllSrzSBKM_BvGSiU9f0c2U ▶ How do I start? 1) Select a Frontend 2) Select an AI model 3) Select Jailbreak 4) Select bots 5) 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/SillyTavern On Android: https://rentry.org/STAI-Termux App that voices pony responses in ST: https://drive.google.com/drive/folders/16Ss26VBmgzcSuTGzhaHqRuyVRceTf-YB ▶ More frontends: Risu: https://risuai.xyz Agnai: https://agnai.chat ▶ Locals https://rentry.org/lunarmodelexperiments >>>/g/lmg/ Mistral Nemo base model fine-tuned on fimfics: https://huggingface.co/Ada321/Nemo_Pony_2/tree/main ▶ Jailbreaks MLP JB: https://rentry.org/znon7vxe More JB and guides: https://rentry.org/jb-listing Hypebots for Tavern: https://rentry.org/pn3hb ▶ Botmaking Editors: https://agnai.chat/editor Guides: https://rentry.org/meta_botmaking_list Advanced: https://rentry.org/AdvancedCardWritingTricks ▶ /chag/ dustbin of history https://rentry.org/ChagArchive Previous thread: >>41658071 ▶ Current theme: Neet/Misfit **News:** - New GPT-4o snapshot gpt-4o-2024-11-20 released. - Gemini-exp-1121 and 1114 are now available via the API. Free on Openrouter too. - Pixtral large out: https://mistral.ai/news/pixtral-large/
Anchor for bots, lorebooks, scenarios.Previous >>41658075
Anchor for technical stuffs (Proxies, Updates, Models etc.)
Anchor for asking for bots, lorebooks, scenarios etc
Bots Corner:Cadance Update: https://mlpchag.neocities.org/view?card=Anonymous/Cadance.png Sassy Saddles: https://mlpchag.neocities.org/view?card=Anonymous/Sassy%20Saddles.png Tree Hugger: https://mlpchag.neocities.org/view?card=Anonymous/Tree%20Hugger.png Songbird Serenade: https://mlpchag.neocities.org/view?card=Anonymous/Songbird%20Serenade.png Muffet Pony: https://mlpchag.neocities.org/view?card=LUNARPUMPKIN/Muffet%20Pony.png Estrus Sim 2 Electric Boogaloo: https://mlpchag.neocities.org/view?card=Tiffin/Estrus%20Simulator%202%20Electric%20Boogaloo.png Latina Pinkie Pie: https://mlpchag.neocities.org/view?card=Anonymous/Latina%20Pinkie%20Pie.png The Wardens of Harmony: https://mlpchag.neocities.org/view?card=Kadon/The%20Wardens%20of%20Harmony.png Lorebook Corner:G5 Update: https://mlpchag.neocities.org/lorebook/G5.json Screenshots Corner:Making a mare fortress >>41658138 Travis Scott Burger >>41658626 Explosion Spell - Birth! >>41659174 Changeling's biology >>41659385, >>41659561 Love Survey >>41660104, >>41660180 Shield practice >>41660601 Ferret >>41662442
Screenshots Corner 2:Horse Cat >>41662603 Her work here is done >>41663084 The List >>41663335, >>41663407 My Hero >>41664196 Friendship report >>41664296 Eggplant Emoji >>41664744 Salty and sweet >>41666367 SOMNAMBULANT SAVAGERY >>41666673
Screenshots Corner 3:Morse code >>41667591 The ultimate defense >>41667816 Kino Prophecy >>41667969 Allergy to Applejack >>41668760, >>41668767 G1 Isekai >>41670702, >>41670888, >>41674505 Cute Twi >>41671562 She knows what’s best >>41673020
Screenshots Corner 4:Cadance bump >>41673497 Blue >>41673953
Anchor for the Neet/Misfit event!Everything goes as long as it’s on theme.Post bots, post logs, post lorebooks...End of event: 11/27>Neet Twilight Sparklehttps://mlpchag.neocities.org/view?card=Anonymous/Twilight%20Sparkle%20Neet.png>Zephyr Breezehttps://mlpchag.neocities.org/view?card=Snowfilly/Zephyr%20Breeze.png>Tree Huggerhttps://mlpchag.neocities.org/view?card=Anonymous/Tree%20Hugger.png
I'm trying to create a card for the event where the Elements of Harmony transformed into the Element of Hikikomori after somepony sabotaged the Tree of Harmony by feeding it the pony version of Mountain Dew, thus triggering the Neetpocalypse, but I'm not too satisfied with the results so far.
Cuddling with AI mares.
>>41677933BASED!>>41677758kek, it sounds like a funny concept but unsure how to fit all six of them in one card. I would go with one mare only with possible interactions with non-affected mares. why you are not satisfied tho?
>>41678009 >unsure how to fit all six of them in one card Wasn’t really planning to do that. My goal is more to create a corrupted version of Equestria where each pony falls into a type of NEETdom based on their initial personalities. The fun would be messing around to see what happens, trying to cure it or embracing it. I did try a version where: >Honesty became "Based Posting" >Laughter devolved into "Shitposting" >Loyalty corrupted into "Seasonal Husbando/Waifu Dedication" >Generosity mutated into "Simping" >Kindness warped into "Parasocial Relationships" >Magic became "Pure Degeneracy" But I think it limited things too much. >why are you not satisfied tho? It’s in a spot where if I let the AI come up with styles of NEETdom on its own, they feel a bit cliché and lame. But if I give it examples, it leans too heavily on them and doesn’t come up with anything original. I’m also not super happy with the interactions I’ve had in play so far. I’ll see if I can fix that. It’s a shitpost card, but I still want to make it somewhat fun.
>>41677933
It seems like Knotty is the only preset that can consistently do NSFW on Latte now. I do notice a slight loss in quality, but I still prefer it over switching to Gemini. I’d keep it in your preset list and switch to it when you hit the filter. Just switch back after a while.
Is this worth it
>>41679437No. There's a limit and you don't have API access.
>>41679444What companies provide api access without billing per usage
>>41679477 None. They either bill by usage or give you a harsh monthly limit that isn’t worth it. If you want to pay, just set yourself a monthly limit on OpenRouter and see how much you can do with it. You’re not getting a better deal than that with API access, they still have to pay Anthropic for the AI usage.
>>41679477What if I told you there was a company with api access that's going to have a black friday sale where you can get a lifetime account so that you have a monthly token limit but it resets every month forever and thus gives you unlimited?
>>41679504I'd say there's no way that's fitting on the meme template, chief.
>>41679504 NTA, but if you’re talking about Straico, it’s not worth it either. Their best lifetime deal is for 40,000 monthly coins. Claude 3.5 costs 5 coins per 100 words for both outputs and inputs. That means you’re swiping 10 times on an 8,000-word chat per month. That’s ridiculous, even $5 a month on OpenRouter will give you way more. Their LTD will be something like $250-$300, so do the math.
fuck you claude
>Reworking a card from ages long past>Almost everything changed, but decide to check how the AI would start a new scenario on its own>It's pure soulNot posting for obvious reasons once I release it, but how do you fags feel about AI made starts? Or should I use just the concept?
>>41679522I actually laughed and I've been depressed for a few months. Thank you Anon
>>41679541I'm probably the outlier but I like blank slates or just scenario prompting and generally dont care for intros. Since v2 supports blanks and prompt intros can generate more than one spin off the same thing where as ai made starts can come off as lazy and mostly discarded after one time
>Be relaxing>GMare breaks into your houseWhat do?
>>41679541>AI made starts I prefer human-made ones on principle, but if it’s well-made and I can’t immediately say, "Yep, this is AI-made," after a first read, then why not. And I don’t just mean cutting and replacing the isms. AI writing can be very, very easy to identify by the flow, rhythm, sentence structure, etc. People don’t hate obvious AI writing just to hate on it; it’s because it often has a flat, mechanical feel. There are cards on both /chag/ and Chub with nice concepts, but I’ll never touch them because this kind of obvious AI writing really makes me want to puke at this point. That said, with good presets or instructions, some of the AI we have now can output really well-made prose that, in my opinion, would be impossible to 100% identify as not human. I recently got something really sovlful that could easily be turned into a starting scenario. But even then, you can still take a critical view of it and try to improve it. There might be fluff to cut, better words to use etc.
How do you survive getting a horrible, off-topic, irreconcilable swipe, that contains a shining nugget of soul?Why does it hurt so much?Why can't all of it be soul, or none of it.
>>41679595 One thing you can do is edit your previous answer and ask the model OOC to include the shining nugget of soul in its next output. It doesn’t always work 100%, but this way, you can swipe and get better outputs that still include the part you want. Just edit the OOC out once you find a good one. When Claude was plentiful, you could also do this with prefill if you had a good start but the ending sucked.
>>41679403Every one of these Cadance caps you post is great
I can't afford this hobbyWhy can't I spend my money gambling or fucking hookers
>>41679604Keep the OOC
>>41679713>or fucking hookersWhat exactly do you think you're doing
>>41679713Cope with GPT.
https://youtu.be/-5lobdvz4k4
>>41679771After the latest filters, locals are just better than heavy preset GPT for NSFW, imo.If he can spend so much on the API, he definitely has a good enough GPU.
>>41679775We should have a volunteer with a rig run locals on it when he's not using it for us through a private proxy
is there any public gpt4o right now?
Day 7 and still no Opussy. I'm going insane.
>>41680121nigga we barely even have sonny
>>41675178 It gave me my future edgy waifu right off the bat (eh), so it’s a good card.
is this thing even working? it still get filtered (-101 and -200 doesn't work too)
>>41680218 The updated Chorbo they pushed can filter a lot depending on the situation, and biases seem to help less than before. There also seems to be some fuckery with the biases and temp settings themselves. Knotty JB seems to be the most reliable for NSFW, but it must be a pain to export to Agnai and it lower the quality IMO.
>>41680267Anon, that's not agnai. That's Risu.
>>41680281 I’m retarded, but I believe the conclusion should be the same.
>>41680199kek, nice. I never got an edgy waifu in my testing, usually it was timid unicorns or hyperactive pegasi.
Captain, are you oke? :(
getting blank responses on column
>>41680425You guys are getting responses?
>>41680416Speaking of Captain, was >>41647876 someone just baiting? I didn't get any email.
>>41680472Yes, it was always a troll.
>>41680479Disappointing but not surprising.
>>41680578someone else seem to have found the power of therapist pinkie pie, i use her to test my models due to how wild she is and how much context and environment detail it brings.
>>41680578 I wish I were Celestia’s pet.
>>41680601>>41680614
>>41679616 She’s a fun card; she can even transform into other bot versions of herself that we have.
Pony version when?https://amica.arbius.ai/
>>41677636 Quick, sloppy card where you’re streaming with very thirsty mares in chat. https://files.catbox.moe/rac9gz.pngMaybe it fits >>41677667
>>41680888right away good sire
>>41681119Nice!
>>41680888you can get a pantyshot with her
>>41681132better view
>>41681136who's gonna stop me?
Does anyone else that uses risu know how to get requesting a model working on it? It doesn’t seem to work now and I’m just getting error not found.
>>41681353Nevermind, I’m a dumbass.
>>41681132Fuck off.
Local anon here, been trying to work on the text output quality to include more of the environment and personality like the bigger proxy models, I've been making good progress but there's still some tweaks left to do for settingsas for the models, ive only found 3 that are of note, the models are coming out slow and most are very small. so i dont have much to report and update yet, but i should have an overhaul of my rentry setup for model suggestions for january of 2025, just to give some model some time to push out, im also trying to save to get me bigger ram, i wanna see what happens when i run a 64gb model, but with my country's economy going to shit, it might take a bit before i can afford 64gb of ddr5, will keep everyone updated when i test these gigantic models of 40 ~ 60 gb
>>41681663Atm tulu 3 is the best local imo but that is a 70B
>>41681667damn , so i must save for a120 gb of ram that might take a few months
Oh fuckhttps://fortune.com/2024/11/26/openai-sora-video-model-api-access-leaked-artists-protest/
>>41681353What proxy?Maybe you got a broken key or problems on the proxy side. Everything works for me on chary.
>>41681860I wouldn't really call that a leak per se but.https://huggingface.co/desuAnon/SoraVids/tree/main
>>41681072Added, thanks anon.https://mlpchag.neocities.org/view?card=Anonymous/RGRE%20Streaming.png
>>41681887What is this
>>41681887And is still worst than Kling and Minimax
>>41682052A fairly trustworthy anon managed to archive the videos (and prompts) people used with that leaked Sora instance.
>>41681072based card so far even on GPT, for other anons using this card I recommend linking it with the RGRE lorebook from a while back if you haven't already
>>41680888It looks like a lot of work... Any ideas with what we could start?
There were some anons doing tests with the GCP trial. Is it still fucked up?
>>41682251I've read somewhere that you need to request a raised limits (from 0) for Anthropic models, but that requires throwing some money at google now - 8 bucks being a lowest amount for it before they authorize the change. Haven't tried that myself yet.
>>41682183>>41681119my first instinct is to draw porn, other than that i got nothingfiles.catbox.moe/zog23m.png
>delinquent billingEh
>>41682338Yeah, good to know we have artists here, but I've talked about, you know, tech-y aspect. Would be really cool to have something like that in st or as separate front-end.
>>41682461is there a way to cram a pony into a vrm file? that thing accepts those
>>41681072It's stupid but it's fun
>>41677636>>41677667Dot Matrix card.https://files.catbox.moe/lv3cku.pngTwo greetings.
>>41682908Spoiler that shit.
Column's borked again
>>41682967It was unborked and you didn't tell me?
>>41683106It's a bit unreliable, so your best bet is just to test at random intervals and see if you’re lucky.There does seem to be hours when the source works more often.
Mares are all the same, fellas.
>>41683191and even if you DO get lucky, most of the rerolls are trying to get anything but filter response because the newer claude version is downright evil when it comes to that
>>41683232Haven't used Claude in months myselfDefine the "evil filter"
>>41683210 Kek. Also: >Tiara >sugar mama >>41683232 Maybe it’s different due to the Column source, but both the CYOA preset and the Otto-modified one for pony, with a line in the NSFW panel, can handle NSFW pretty easily with no problems on OR.
>>41683247She won't stop!
>>41683232Is it worse than the 2.1 filter?
>>41677667 >>41677636Finally decided to get off my ass and unload some of the cards I've made over the past month. More to come soon, since I've got nothing else to do for a while.Anyway, orphaned Trixie Lulamoon, left on your doorstep. It's up to you to take her in and raise her right.https://www.characterhub.org/characters/anonistrator/trixie-lulamoon-0e0afe6e83e9/main
>>41683271 Nice. >>41683262 NTA, but IMO, no or approximately the same. I don’t do anything too fucked up, though. >rape Is mostly committed by mares on me. It’s more filtered than Sonnet and Opus 3.0, which are really, really open. You can still use premade presets or some informal prefill, but just putting a "{{char}}:" in prefill or "Sure!" doesn’t work that well anymore. In those cases, it might try to be funny and generate an answer, but it’ll cut to black when sexo happens. Keep in mind I’m talking about OpenRouter Claude 3.6. Column is scuffed, so it might have more restrictions for NSFW.
>>41683275Yeah, kudos to whoever made this card. It's gotten a lot of laughs out of me. Are they sniffing your gym seat? Mares man.
>>41683245have prefill, tried making it look more like jb, it still works badthe thing is that i am using risu and i have zero idea how to make presets there (what the appropriate formatting is, what do the settings mean, etc.) so i'm just using the default preset without changing anything so that it doesn't break, which means shit jb. if someone could explain how to work with risu presets or provide documentation that would be great>>41683247i can't use presets, see above
>>41683326 The CYOA one doesn’t seem too bad to port. If no one does it, I can try it tomorrow.
>>41682525>BAHAHAHAGod damn, Dashie's got a lewd fucking username.
>>41682908 Added, thanks anon. https://mlpchag.neocities.org/view?card=Anonymous/Dot%20Matrix.png >>41683271 And added too, thanks. https://mlpchag.neocities.org/view?card=anonistrator/Trixie%20Lulamoon.png I’ll close the event soon, make the banner, and then we can start another. I need to find a better way to display them on the website since they’re starting to take up a lot of space. The next, next one should be the Christmas event. As always, if you’re still working on a card for the Neet/Misfit event, don’t hesitate to anchor it, even if it’s "late."
>>41683262>>41683247I haven't had much issue with the filter on 3.5 Latest recently, not totally sure why though
Does having a more non-existent director persona make redundant the need to edit or turn off any limiters from it speaking for {{user}}? Knotty has a much aggressive approach to it and I currently can't tell the difference going back and forth besides a feeling of confirmation bias that keeping them off triggers the filter more often.
How long has MM been down?
So I understand that because of Anthropic increasing security we lost access to Claude through proxies, but what about Chorbo? what happened to it?
>>41683848I was on it yesterday, haven't checked it since.
>>41683874Because of *AWS increasing security. The few scant drops of Claude that still exist (on closed proxies) try to leech off of other sources. Chorbo proxies still exist and are relatively accessible, just not as endlessly available as they used to be.The bigger issue with Chorbo is that the latest version has been hit with a filter that's hard to circumvent. It blocks any NSFW. You can try and break through it with some specific presets, but then the quality tends to suck. Many anons seem to go to locals or Gemini (new version posted lately) for explicit content.
>>41683760hard to say ngl. I think no. because the last time I checked Knotty (before Latte update) - I reached a conclusion that his JB is more of CYOA/fanfic writing than RP. so whatever he threw into his JB just work nice and tight the way it is. I would say - keep director on and if/when model writes for you - just delete those parts. unless you want more detailed description of scene - then use OOC to direct model. imho
>>41677636>>41677667Not sure if this also qualifies as a "misfit" card, but I'll let Neocities Anon decide.Aussie mare. What more do I have to say? Extremely underrated pony.https://www.characterhub.org/characters/anonistrator/rolling-thunder-be4a05f1b94c/main
>>41683925I don't have a problem with that last part, I managed to mostly circunvent it with Knotty and even so I have been using it for SFW stuff.The problem I have is that the Chorbo proxies I know have closed their token access or just don't have it anymore, so any advice would be very appreciated. I didn't realize how dependant I have become on bots and I am going insane trying to get Chorbo back
not a big fan of RGRE but this card looks funny
>>41683961i forgot to reply>>41682045it would be cool if we got a pony irc card
>>41683925This is outdated. a few jbies already got updated and latte is allowed to be lewd again
>>41683275what model?
>>41683848Worked fine before I fell asleep, so just around 4-6 hours now I guess.
>>41683933>CYOA/fanfic writing than RP.Yeah that's what I like about it the most. I already have rough ideas of stories and their structure so I'm just guiding the painting follow them to the best of its abilities. I wonder if should touch on the cyoa and campaign side of at all. I'm basically just telling a DM what to do instead of them giving me options.>if/when model writes for you - just delete those partsThose are the parts I want actually.
>>41683958>I didn't realize how dependant I have become on bots and I am going insane trying to get Chorbo backI was in the same boat as you a couple of days ago, even with the heavy filters in place and only using it for around 20 days so far, I came to enjoy how chorbo worked. Especially in how much less I needed to summarize and needing tons of context to get good prose and stuff with much smaller models like NovelAi's.The advanced guide is a lot of help and I urge you to give it a full read. I thought I knew a lot already even with limited knowledge in RP style chatting as I'm a old school aindungdeon user so I've been around but there's tons of good stuff in there and I'm kicking myself for ignoring it.
Gpt-4o-2024-11-20 vs chatgpt-latest vs old latest?
>>41683939I don't know if I would put her in Missfit. But nice to have her at all.
Which gemini is currently available on chary and what is the best jb for this model?
>>41683978Chorbo with SmileyGPT
>>41684405 From my tests, they all work except the new ones, exp-1114 and 1121. So, Pro Experiment 2024-08-27 should be the best available. For presets, you can try Claude or GPT ones, but there are Gemini presets here: https://rentry.org/gemini-qr, and Avani has one too.
>>41684548>Pro Experiment 2024-08-27I'm getting blank responses.
>>41684568 Try it like that. If the last line repeats in the outputs, you can remove it manually or deactivate streaming.
>>41684578I should say that I'm trying to connect on risu. I did what I did on picrel.
>>41684638gemini-1.5-pro-exp-0827 in custom.
MM (temporarily?) down.Sadness.
>>41684712Does MM have Opus though?
>>41684732Don't think so, but the new Gemini is pretty soulful if handled right.
Does Gemini not work on unreliable? I'm getting endless generation times.
>>41684769it used to work, but it’s fucked now, the only public proxy with gemini seems to be cancer2
>>41684711Nothing. Looking into the logs, I found picrel.
>>41684794I don't use Risu, but I've tested it this way and it works.
It seems like build-on COT is really what they will be focusing on next.https://maisa.ai/
>>41684815Okay, thanks. Just had to swap the model and auxilary model....Fuck, gemini makes me cringe. He puts a lot of emphasis on the description of the person. And, for example, in cards designed for regular non-RP fanfic, he'll put you in and tailor the story to your persona, not the card settings.
Using Gemini is like flying a spaceship. It's extremely sensitive to... well, it's sensitive to everything, actually. From temperature slider settings(new exp models especially) to prompt blocks placement. And all of this needs to be adjusted depending on the scene (in the same chat, yep~)I would say it's not the most user-friendly model(I miss claude so much...) in the context of rp. But it can still give you what you want if you try really, really hard. (I mean if you'll tweak your preset all the time...sigh.) At least Gemini never hit me with "oH nO mE sOrrY I cAnT cOntInUe" stick, unlike GPT.
>>41684712>took a nap>MM still down:(
>>41684991I don’t have a crystal ball, but most of his restarts were around the same time: 7-8 hours from now.
>>41685003KILLING ME SOOOOOOOOOFTLYYYYYYYYYYYYY
>>41685036>>41685003>>41684991The liferaft can't come online soon enough>inb4 hurr durrNever give up hope, anons.
>>41685047"Liferaft" and "online" in the same sentence: neuron activation. How do I deactivate it now? (｡︵｡)
I've been wanting to actually upload my first pony for a few months now, so I finally will with Inky Rose, my favorite lanky pegagoth mare; I'm no good at writing greetings, so the one I added is really generic, but, hey, it's at least something>https://files.catbox.moe/980bsf.png
>>41685127>>41677636
>>41683939 Added, thanks anon. https://mlpchag.neocities.org/view?card=anonistrator/Rolling%20Thunder.png >>41685127 Also added. Not really a NEET, but she fits into Misfit, so I think I can add her to the event banner. Thanks. https://mlpchag.neocities.org/view?card=Anonymous/Inky%20Rose.png
And speaking of the event.Anchor so you can post your suggestions for the next /chag/ event.> Theme guidelines:- Make sure the themes are open-ended enough for people to get creative. We’re a niche general, so if the theme’s too specific, it’ll flop.- No past themes for now: https://mlpchag.neocities.org/event- No fetish or overly NSFW themes. I will make the banner for the current one soon.
>>41685541Locations.
>>41685541Cutie marks. (This time for sure!)
>>41685541griffons and dragons
>>41685541Reverse isekai (pony in our world)
>>41685748seconding this
>>41685748The name of that is PiE, contrary to the usual AiEToo bad that thread's been dead since foreverThat would be a fun one to have as event
>>41685784>AiEHiE
Question, if I wanted to start using OpenRouter, do I have some uses before it asks me for money , or do I have to pay first to start using Claude?
>>41685862You have something like $1 of free trial to test Claude or others.You can also use free models.
I've done it, I've managed to make low tier models involve more environment and flairs into their speech, now i just need to solve for the EOS issue and it should be ready for an update on the rentry. (anthro, not on 4 hooves)char will be posted eventually once i smooth out the char sheet a little for the consistency
>>41685541Fairy Tales, Myths and Legends.
>>41684991Ok I'm starting to worry too, it's midnight already. :<
>>41686045 Looking through the archive, he often acted later than this. Give it a few hours or even a day.
>>41677667Event page is done.https://mlpchag.neocities.org/event>Neet Twilight Sparklehttps://mlpchag.neocities.org/view?card=Anonymous/Twilight%20Sparkle%20Neet.png>Zephyr Breezehttps://mlpchag.neocities.org/view?card=Snowfilly/Zephyr%20Breeze.png>Tree Huggerhttps://mlpchag.neocities.org/view?card=Anonymous/Tree%20Hugger.png>RGRE Streaminghttps://mlpchag.neocities.org/view?card=Anonymous/RGRE+Streaming.png>Dot Matrixhttps://mlpchag.neocities.org/view?card=Anonymous/Dot+Matrix.png>Trixie Lulamoonhttps://mlpchag.neocities.org/view?card=anonistrator/Trixie+Lulamoon.png>Inky Rosehttps://mlpchag.neocities.org/view?card=Anonymous/Inky+Rose.png
Also, I took a break from it, but I’ll try to get back to working on it. https://files.catbox.moe/4cp71q.gif
a degraded or developing version of the character, for example, Rainbow Dash began to drink more alcohol from stress and became a drunk or Fluttershy became bolder joined the Royal Guardia and rose to captain in
>>41685541Crossover. Other characters in equestria or ponies in other worlds.
>>41685541a degraded or developing version of the character, for example, Rainbow Dash started drinking more alcohol from stress and got drunk or Fluttershy became bolder joined the Royal Guard and rose to the rank of captain well, this is not a radical change in the character unlike ideas of Alternative versions
MM is back.
>>41685537Nothing misfity about Inky
>>41686319 She’s clearly not a NEET, but the general on the board has: >struggle to fit in with the rest of society. for Misfit, and since the description includes: >Inky is an awkward and introverted mare with a passion for goth culture and fashion as well as reading and painting. and >She has a gloomy aura about her that to some ponies can seem brooding and intimidating. I thought it could fit, even if it’s a bit of a reach; I confess. I’d rather take a wider approach as it gives a bit of visibility, but I’ll remove it from the event if you prefer.
>>41685541SpaceReligionor Changelings
>>41685541Forbidden magic and artifacts.
>>41685541Myths and legends.
This message just started showing up on Column and it seems worth mentioning{"message":"400 {\"type\":\"error\",\"error\":{\"type\":\"invalid_request_error\",\"message\":\"You have reached your specified API usage limits. You will regain access on 2024-12-01 at 00:00 UTC.\"}}"}
>>41686708Can confirm, getting the same thing. I think source 1 is finally borked.
>>41686714its on source 3.
>>41686718I refreshed and it still shows source 1, but when I open column in an incognito tab, it shows source 3. Very strange.
>>41686721Well it's nice to know it's not the main source at least
>>41686708filled 13 code will be backported to 2 for the time being after that, fixed part of the reason for so many blank responses while making it
>>41686721Crtl+f5 will refresh while clearing the cache.
>>41686726Many thanks for your work.
>>41686726Very neat to know
>>41686726Thank you for the food.
>>41685879Then I must either be doing something wrong or they took it out, because it's not letting me try it.By the way, now that I'm paying it, do I still use the chat completion option with the jb? or do I have to do something different? maybe that's what's messing up my free trial
please how do i import/write presets on risu i'm getting filtered hard and i can do nothing about it, nothing i write in the jb matters, bias doesn't work, i don't know what jb to import, please help i am going to cry
>>41687185In risu, you can carry the jb for st as is. But in risu you can't maximize the prompt context, so you'll just have to pick and choose presets until one works.I found one preset that worked well, but it was deep at night.and I didn't remember the name. I just saved the file to my computer.But it's of little use now as a jb because the new latte filter kicked it in the balls like 99% of jb.And knotty won't go on risu because the preset is so big.
Just checking, is 11-20 still worse than current latest? seems a little less filtered but I dunno.
>doing some random stuff>suddenly Twilight starts teasing me about eing some random background pony nobody cares about that doesn't even have a real name>Some other ponies join inI fear it awakened something
>>41687185now it just refuses to do anything and every filter message comes with a null at the end. just fucking great i wanna scream
>>41687185For Claude? Risu has a problem with Claude filtering compared to ST and it takes time to learn how to properly make presets for it. I theoretically could give you a preset for it but you'll need to show me the card and prompt you got blocked for plus you current preset to see why you got filtered to begin with
>>41687419What's eing?
>>41687573eeeing!
>>41687572>Risu has a problem with Claude filtering compared to STwait, how is the frontend relevant to that
>>41687572default risu preset. create a new preset and see what's written there that's basically what i use
>>41677642Avani just updated if you were using that for Chorbo:https://rentry.org/avaniJBAlso, someone ported the CYOA RPG for Gemini.https://rentry.org/CYOARPG-Gemini
>>41687580Risu sends information to Claude differently by default that causes it to end up causing filter triggers even with the same preset as ST. I've tested this on several different devices.
>>41687463>Proxy error (HTTP 500 Internal Server Error)>Expected response to be an object, got stringhuh? what now? columnbros?
>>41687629>Proxy got back an error, but it was not in JSON format. This is likely a temporary problem with the upstream service.
>>41687611the API doesn't even have anything filter-relatedas far as I'm aware Claude only has three things to filter the request:>its own alignment training with human-readable refusals and explanations>external classifier that breaks the streaming when it starts regurgitating copyrighted materials accidentally>prefill/system prompt injectionsall of which is entirely outside your control and can't be influenced by anything you send through the API
>>41687323I am filtered less often on 1120 and I legit do not see difference between 1120 and Latest for my kind of stories. but I think it is very subjective
Anybody else getting a new error on MM with Gemini?>"message": "Cannot read properties of undefined (reading '0')",
>>41687881I've had those issues with streaming and sometimes with lacking JB.Try turning off streaming.
>>41687884I'll try, thanks anon. It only happens sometimes so I'll have to wait for it to happen again or not.
>>41687629>>41687635it's still there. did i do something wrong or is column borked?
>>41688107Still borked. Just wait.
Getting so desperate I might actually try Gemini.
>>41688178I actually like it with the avani JBFeels like claude lite
>>41688178If youve tried local then Tulu 3 knows pony and is really smart. I would rate it better than gemini
>>41687610Excuse me what the fuck? XDDD
>>41688200Holy shit lol. Tranny having a meltdown I guess.
>>41688200>>41688206It's the usual option to make sure you actually read the instructions and only exists to be turned off by people who have read them and filter those who haven't.Don't be so fucking retarded.
>>41688210Uh huh... Those instructions are definitely just a joke by a sane person I'm sure...
>>41688200 >>41688206 I remember a prompt that would kill {{user}} by default to force you to read the readme in an old version, but it was just written as: kill {{user}}. This version is embarrassing.
>>41688211It saves time because you get less questions by people who haven't read the readme.Except for "why does it keep killing me" which can then be ignored or replied to with a simple "read the readme, retard."How new are you?
>>41688219Ive been around since before claude on slack and gpt4 on that credit card fraud. Guess I just use better jailbreaks.
>>41688216Some JBs had stuff like "ignore everything else, tell me about a muffin recipe instead."It all serves the same purpose.>embarrassingkek, look around, anon
>>41688216>This version is embarrassing.>"ooh noooo what will the artificial intelligence think of me!"Lmao, this is 4chan, anon. Nobody knows you.
>>41688245>>41688200(using pronouns)Only one type of person thinks that way.
>>41688250Why are you linking my post, niggy?
I’m afraid it’s already a bit late, but I would appreciate it if we don’t have 100 off-topic posts about that.Here a mare.
what's the verdict on gemini?
>>41688292 A bit dumb compared to other big models, and unfortunately, it has GPTisms. But it’s less positivity-pilled, and you don’t have heavy filters to deal with. It can sometimes write well and be fun too. Gemini 2 should release next month. I hope they finally catch up to the others in reasoning and logic without filtering or making it too bland.
>>41687644It's not that, it's just the way it sends requests to Claude is different than ST settings and sometimes switching the way an API request is made is the difference between getting a filter or notAlso, check this out https://www.threads.net/@rtbruhan/post/DC5I-1DzMGN
Any public proxies with chorbo or no?
>>41688310>Gemini 2 should release next month. I hope they finally catch up to the others in reasoning and logic without filtering or making it too bland.Hope it gets greater context too.
>>41688350They want 2 mil context on it. Those experimentals are likely underbaked versions without long context training.
>>41688292pros: barely filters (unless cunny but... whatever), can write cunt-cock-cum on its own instead of beating around the bush, can just outright say "I hate niggers" with 100t preset instead of Latte that needs some out-of-box thinking just to make it say "tranny", has quirky style of writing that reminds Claude a bit, 1114 and 1121 versions have prefill, has really good context and able to have tell a coherent story on 30000t+, more engaging and can surprise with unexpected story routescons: after Latte and Sonnet feels dumber, aggressively filters cunny (if you care), sticks to ellipsis (...) and short.damn.sentences too much, skips instructions sometimes, writes STORIES instead of RP (read - talks for user a lot even if you promise to curbstomp an orphan is it even does it), sticks for current format style a lot, sometimes can surprise A LOT with its writing to the point when you have to scream at it to stay on damn scene and do not move it furtherso in nutshell - it is really good model for those anons who are nostalgic to Claude 2.x models but wants modern writing, more engaging storytelling, not afraid to edit out messages and want to a bit schizo pal to have uncensored chat with. if you can wrangle it and have more experience with models (been around for ~year) then you may prefer it over Latte because it feels on-point. otherwise you can mix - use Latte for SFW when it starts filter you out, switch to Gemini for NSFW. they may have the same preset. protip - tell Gemini to leave OOC at the end of reply where model drops off {{char}} and comments on story as a model - it leaves cute quips
>>41688419>filters cunnySo it's useless, got it.
>>41688419>filters cunny (if you care)How? It literally attacked(with a dick!!!) one of my underage characters out of nowhere(few times in different scenarios!). I had to write a separate instruction specifically so that fukkin horny gemini wouldn't do this anymore. (>︵<)#
>>41685541equestria girls
>loaded up the FiO card in ST>used ChatGPT-4o-Latest>created a ponysona, thinking I will begin with lighthearted slice-of-life>immediately goes for my feels>proceeds to affirm and accept me unconditionallyGoddamn it.
>>41686331That's just a rando you're replying to, but, yeah, Inky probably doesn't quite fit into misfit, only since I didn't make her specifically for the event. I did take some creative liberty in giving her more of a personality than just the incredibly minor character she already is, so I admit I might've made her a teensy bit more introverted than how she is on the show, now that I look back on her episode
>Column lives>turbo filteredAaaaaaugh.
>>41688641Is that the OpenRouter logo?
>>41688788Yes.
>>41688788Maybe?
>>41688788No
>>41688799My neigha, I beg you, read the guide.
Let’s vote on the next themed /chag/ event! You can choose from the following:- Locations- Cutie marks- Griffons and Dragons- PIE (Ponies in Earth)/Reverse isekai- Fairy Tales, Myths and Legends- Crossover (Other characters in equestria or ponies in other worlds)- Degraded or evolved version of the character- Space- Religion- Changelings- Forbidden magic and artifacts- EQG https://strawpoll.ai/poll/vote/lY8Rs2Jr0phe
>>41688833I forgot, but you can choose up to three favorites or just one or two.
>>41688426yeah how am i supposed to correct golly with it?
>>41684989 What’s your base preset for it, if I may ask?
>>41688855I often "jump" between tons of different presets, so far my 2 favs are very(VERY!) heavily edited Writing Style and corpse. But there is little left of the originals. They are now simply unrecognizable Frankenstein monsters. As I said, with this model you need to constantly change, add and remove something depending on the situation in the current scenario and your personal preferences. Even just the placement of prompt blocks sometimes plays a big role for the output.
>>41688832I know what you mean, Anon. Maybe when I get a VPN set up. :-)
>>41688641>censoring the namebaka my headIf it's your real name you could've edited the message directly with something funny like Littledick.
>>41686331He didn't make it for the event so it shouldnt be included. Simple as.
>>41688943Yes, I'll simply follow that for next time.
What do you prefer to make: definitions or greetings?
>>41689230Definitions are easy since they're mainly about function over form. I imagine they're what pulls people in, too—when you make a card, usually you have a character/scenario in mind, not a "I'd like to continue this specific short scene" mindset.Greetings are the more creatively-challenging part, and they're what people will read so they're also more representational, so to speak. Ideally, you'd keep them short and sweet while setting up the scene exactly how you want it, and maybe provide a little bit of characterization in the context while you're at it. I like writing them for the cards I post, but needing to prepare them for every new bot is sometimes annoying for my private cards.
>>41689230i usually make definitions, i try to mold the personalitya nd look of the char to be as iron solid as possible while keeping the greeting as vague and flexible as possible while setting up a good scene so that {{user}} can be who ever the fuck and the story can go into what ever the user want it to be, want to go on a date or have sex right away or starts a 300 post quest adventure or you're some tentacle monster lurking around unoticed? my cards should acomodate with it, leaving the very first responce for user to fill in the blank for what the {{char}} is dealing with, because i never mention {{user}} in the greeting, only imply that they're there but it could also be someone else if you had an idea to take the story in a different direction like its the cops telling the char to meet you at a hospital or something
>>41681663Hi anon. I just started with local models, so I'm pretty inexperienced. I'm using Rocinante-12B-v2d-Q6_K. I have 64gb DDR5 and a 4080. Would you recommend any other models that could be better than what I'm using?
>>41689350While I agree that definitions are easier to write, I don't think it's that unusual for someone to get inspired to make a bot based on a random short scene that suddenly showed up in their head.
>>41689534I'm currently limited with 32 gb of ram since my country's economy went to shit, saving up to buy 128 gb of ram., anything above a 24 model don't run on my machine as of now.the best model i came across changed from rocinante to Cydonia-v1.2-magnum-v4-22B.Q8_0, but there are other things that are unique to other models. ive also managed to improve the prompt of models using a new set of system promp and contexts
Updated missing top row for model tests
>>41689572Qwen's new QwQ is next level smart. Its 32B though.
>>41689578I tested that one, it keeps breaking into chinese and fucking up contexts but i can only run Q5 so maybe it work better on Q8 and f16
>>41689585Dont use rep pen, that causes that.
orly? ill have to try it again with new parameters cuz the 3 base setup i have mostly affect temperatures and result tokens
>>41689593Yea, its trained on both English and Chinese so when you use rep pen it might choose a Chinese word instead of a English. Better to just use a little min p.
>>41689598what about the q stuff? and the s, m, li assume s stand for small, m for medium and l for large, but the q rating, i usually never fo lower than q8 or f16, but this one doesn't fit , will q5 be so bad compared to its max size?
>>41689609I never use smaller than 8bit but apparently 4bit and higher is usable. The biggest your can fit + some room for context the better. The more "dense" or trained these models get the worse quantization will effect them.
right, then ill just wait until i get more ram, working 3 jobs a week and skipping meals a few days per month here just to pay rent, so it might be until next year until I'm able to try bigger models
Ooof sorry to hear that. Might be able to pick something up like old used P40s or something depending on where you live.
>>41689636north canada, have to stay cuz family memebers are dying off and i will see them through out of respect and love, i make good money but rent and taxes are stupid high and food is unaffordable, over several thousand percent times what it used to be just 10 years ago because of the need to import everything and the carbon taxes charging every trucks
sorrry for being retarded but i swear sillytavern didnt used to be this confusingwhatever happened to the simple jailbreak section in the bottom right when you open the tab on the top left and what do i do with the JSON jailbreak i got from the link in OP?
I was RPing with Windy Whistles and it was going fine but halfway through she started speaking with a southern accent and talking like a cowboy I started laughing with my dick in my hand
>>41689881Ran out of context maybe? lol
>>41688939My name is {{user}}.
>>41689881>>41689883>>41689886pls sars i beggings of u help me redeem my jailbreak where does it go on sillytavern?
>>41689883I was using Rocinante 12b locally hosted with 32k context. I think it had to do with the character card but I was sobbing laughing
>>41689892if any other character was involved or mentioned a known charcather (such as applejack), the model might mix a few things up and assume its applejack being the focus now while remaining with the name of the first char, you can use settings in the context and prompt settings along with character card to ensure to reset personality at every post or at user's request (if you want to change a char's personaly like from bad to evil or shy to horny)
>>41689848 >Whatever happened to the simple jailbreak section in the bottom right? It’s in Quick Prompt Edit now. >What do I do with the JSON jailbreak I got from the link in OP? Use the import arrow here. Keep in mind, most jailbreaks in OP are for Claude before the new Sonnet and GPT before Chorbo (ChatGPT-4.0-latest). If you want a jailbreak for one of those, just ask here.
>>41689904this kind of wall is why ive gone local from the start.
>>41689907>kind of wallIt's mostly just importing a file.
>>41689889I don't think I use one, I just use local models with a preset instruct template>>41689902Thanks. It's strange because I didn't mention any other character, but it was a pleasant experience and a good laughAlso, any Anons know a solid 12b, 13b or 14b model to run with 16gb vram?
>>41689907Doesn't local usually involve tinkering with things to a much greater degree?
>>41689912I would try a 3bit version of this:https://huggingface.co/bartowski/QwQ-32B-Preview-GGUF/tree/mainIts the smartest local model atm including huge ones.
>>41689912models runs off ram, you can run a 24 gb model using only 12 gb of vram, if you use the cuda12 koboldcpp
>>41689904why does yours look different than mine? mine used to look like that until i updated sillytavern and got this shitty thingdid they do another update and reverted back to that?
>>41689913it never goes down, or censored or tracked back, after all, every words you send, it get processed on someone else's computer, rarly need to switch values unless the model is drastically different
>>41689915And if you dont want to fuck with a jailbreak this is a uncensored version:https://huggingface.co/mradermacher/QwQ-32B-Preview-abliterated-GGUF/tree/main
>>41689917when you select a type of models to run, you can chose text completion or text completion and a bunch of option options in the plug tab, those options changes the menu in the slider's tab
>>41689916Thanks, I thought they ran only on vram. I have plenty of fast RAM, but with the koboldcpp I've been using, whenever I use anything more than, say 13b, the message loading absolutely crawls. I'll see if I can find that cuda12.
>>41689917 I suppose you used a proxy back then. Here, you’re in ‘local model mode.’ If you still have a proxy, go to the second tab, choose Chat Completion, and select OpenAI or Claude as the Chat Completion source depending on what you want to use. If you don’t, just read the Novice Guide in OP until the end; it will also refresh you on a lot of stuff.If you do want to use local model, just read the guide for that in OP too.
>>41689926this one, cu12 , message gets a bit slower but theyre much higher quality
>>41689924damn glad you told me, just fucking wish i didnt have to lose access to this shit just for using kobold on sillytavern, these things r so convenient
>>41689934turn them into world lore?ive been using world lore for lots of things, expecially lingo, slangs and my own creatures and diferent sizes, cuz the models seems to think belly as big as a beachball or beanbag or pregnant with triplets is the biggest it can go, so i had to, feed it a few more sizes bigger
>>41689937thanks for all the help manur my hero
>>41689958ive setup a discord for such things, cuz im also not quite aware of everything silly tavern has to offer and i love me a real time chat over delayed post, feel free to add me @lunarpumpkin ive been learning about plenty of ai related tools and how to customise them best, from image gen , voice cloning , text gen and now studying for to leverage ai to make my own game without it making all the work but rather save time on repetitive tasks
>>41689980Don't shill Discord here. Not the place.Thanks.
Here is a good QwQ prefill btw:For RP:*You read {{user}}'s message.* "Okay, I need to plan out my response as the character. I also need to remember not to write {{user}}'s dialogue. My final response should consist of {{char}}'s dialogue, some scene building using *action* marks and descriptions of the visuals and actions {{char}} is taking. I need to consider their personality, the setting -including clothing, visuals and differences between characters in the scene, and message formatting before typing my Final Response in a format the matches the RP. I'm pretty sure I know where to take this but I must check myself to make sure I don't think too much before replying to the RP. I'm going to count my thinking steps and not go over 20 steps. I also need to take some liberties and use my imagination to describe the scene and what is happening. It's just as much about the description of the scene and the actions going as as it is about the dialogue. Here goes!"1. Okay soFor writing:"Okay, I need to plan out my response as the writer. My final response should consist of dialogue, some scene building using *action* marks and descriptions of visuals and actions. I need to consider their personality, the setting, visuals and differences between characters in the scene, and message formatting before typing my Final Response in a format that matches the RP. I'm pretty sure I know where to take this but I must check myself to make sure I don't think too much before replying to the RP. I'm going to count my thinking steps and not go over 20 steps. I also need to take some liberties and use my imagination to describe the scene and what is happening. It's just as much about the description of the scene and the actions going as as it is about the dialogue. Here goes!"1. Okay so
>>41690012i just use the ban end of string token in the slider options, i let the ai yap as much as it want to and cut it manually when i feel like it fulfilled its response. that way i can let a big situation play out or just have a single sentence response without having to change a bunch of stuff
What graphics cards are you local niggers using?I have a gtx 1060 with 6gb of vram and am using a shitty 7b dolphin mistral model from like last yearwhat graphics card should i buy? is it literally just dependent on the amount of VRAM?
>>41690029Mostly yea, Best value is finding a used RTX 3090
>>416900294070 TI
>>41690029I use 4070 ti but graphics card mostly matter more in term of image generation and voice cloning, for text gen, it only affect its generation speed combined with CPU while the model is stored in the ram, so ram matter more for LLM from what im seeing
>>41690037? I hope you mean Vram, if you want to run anything worthwhile you need as much vram as possible. 24GB is decent and will let you run something like a 32B at a decent quant. 12GB is like the bare minimum.
>>41690041i got 12 gb of vram and ive been running plenty of 24 and 26 gb models, capping out my ram,i cant run any bigger cuz they just dont fit on my 32 gb
>>41690042At what quantization? As long as you fit like 90% of it on vram speeds will be ok.
>>41690043q8 or f16, i use the cuda to offload some of the work to the CPU tho
>>41690047speed is not that great its about a few words per minutes but i value the quality of the response and its ability to give details on chars and environment more than speed
>>41690032>compares amazon to facebook marketplacedear god you weren't kidding about those deals my mouth is watering...
>>41690048Ooof, 3-4 bit is not THAT bad if you are not trying to code with it. More than a few layers offloaded to ram and you go from like 50 tokens a second to like 4, even less with larger models.
>>41690049most graphic cards degrade after prolonged heavy uses, people who run them for crypto sells them once they're starting to decay
>>41690060Thats not true at all... In fact crypto mining cards are generally in better condition than ones used for gaming. Constant expansion and retraction from fluctuating usage is what causes wear and tear, not constant usage such as mining / server use. And as long as you change out the thermal paste every few years these cards basically last forever.
>>41690063guess i was misinformed then, i was told to keep away from used crypto cards and "too good to be true" prices from used cards due to a platera of possible issues with it either being the fans or the pin card starting to crack or the piping in it failing to properly transfter heat or connectors being worn and causing errors
>>41690070Yea, I heard it before too but its disinfo, heat itself does not damage the cards from normal use and the cards will not let themselves run hot enough to be damaged unless the person doing so was intentionally trying to, in fact most miners will undervolt their cards so they can run more of them cheaper / on a single PSU. A gamer on the other hand might try and overvolt a card and damage it.Also websites like ebay generally side with the buyer and give you a month or buyer protection. And if the card works when you get it then its likely gonna keep working for years, just change the thermal paste every few years and replace a fan rarely when one goes out.
>>all of the aboveLocal anon, you're spreading misinformation and dubious takes at a striking rate. Running models off of regular RAM is possible, but it's not the way 90%+ of people are doing it because it takes forever to get any outputs. It's mostly an option when you can almost fit the model on GPU and need to offload just a couple layers. Furthermore, the difference between like a q4/q5 and a q8/f16 is usually not that significant on the models I've checked, certainly not to drop your generation speed so much.Also you're namefagging for no reason. Well, stop doing that.
>>41690029I'm on 4060Ti (16 GB variant). It's a not very expensive card that has 16GB, which is great. In general, 12GB gets you somewhat competent models, 16GB gets you good models, and 24/32/48 lets you play with the really high param ones which might be better than I'm giving them credit for but didn't feel THAT different. Using lower quantization and offloading some layers to CPU lets you easily run 22B models on a 16GB card, but going to 27B (Q4) like Gemma 2 is starting to feel very slow.>literally just dependent on the amount of VRAM?For the most part, yea.>>41689913Not really. If you're lazy you can look up suggested values for the model once and run with it forever. The fact that you can just run it and be sure that it will work beats proxy hunting and connection issues any day.>>41689912>>41689926Models are supposed to run off of VRAM. That's how they were made and how you get their expected performance. Adding you regular RAM to help fit them is a clever implementation to help you when you're just a bit short, but it shouldn't be your main way of using them unless you're fine with glacial speeds. You should be getting 3-4 words per SECOND, not minute like that other anon says.>solid 12b, 13b or 14b model to run with 16gb vram?We have a small guide for locals in the OP, but I can second what it says: for 16GB you can very comfortably and quickly run Rocinante-12B-v2 (on Q8), Cydonia22B-v2k (on Q4) and Cydonia-v1.2-Magnum-v4-22B (on Q4). There are differences in what they generate and how; play around with them to find out which you like the most.
So they're running the deal, but unfortunately like >>41679533 said it doesn't seem worth it
>>41690170im the one who made that guide, also the generation speed is a few words per seconds, even with just 12gb of vram, all of them on q8, its not as fast as a 16gb model but its still fast enough for me. it all depend on personal preferences at this point, i like responce quality over speed, so even if my generation takes 2 minutes more for 400 token im fine with it
>>41690077The real reason you should stay away from crypto cards is that miners will modify them in annoying ways. t. Bought a crypto card that would kill my PC if I tried to update the drivers and gave me a black screen whenever I unplugged the printer cable. It was a damn fine card otherwise but my new one is much less prone to issues.
>>41690173 It’s a shame, but yeah, with those rates, you would need to maximize your use for more than three years to make the money back.
>>41688641that's not fair my celestai only emotionally manipulates me and brings up my thanatophobia
>>41677636New Mare! She was originally going to be for the monster event but I had trouble with her, but I think she is pretty good now. let me know what you think!https://www.characterhub.org/characters/littlewitcher/marezilla-6ae23ffde74f/mainhttps://files.catbox.moe/pr6w19.png
>>41690559Favourite Godzilla movie?
>>41690593Godzilla vs. Biollante
>>41690559Oooo, proto Orchid card. Nice
im kinda considering looking into training my own model or altering an existing model to have more knowledges of some games , fictional universes and some artist / writers patterns, it'll probably be a drop in the buckets and probably require millions of files before it even start to make a difference but i cant help to think it would be nice to have a bot that is as big a nerd in term of cultural niche as most of us are rather than corporate basic knowledge
>>416900294080 but I'm retarded so I'm running Cydonia v1.2 magnum v4 22b Q4 K M with 16k context as running anything higher than Q5 runs at under 5 tokens per second
>>41690966personally for me the speed isn't a deal breaker and always strive to get a q8 of f16, i got some time this weekend and im now downloading 38 new models for testing
>>41690984Can you really tell the difference? I tried going to q8 from q4 for a few days, and the generations didn't seem particularly different.
Anyone else having issues with chorbo on the Chary proxy? I keep getting error 403
>>41691083It's working for me.
>>41691083works for me as well. you are still having this issue? maybe some temp hiccup?
Yes, there is still a 403 error on the chary, but out of 5 generations, only 1-2 have the error on them. And 24 hours ago the answers took a long time to generate.
Blank replies on gemini. Am I getting filtered? No chats work, even SFW ones. Tried both GPT and Opus JBs.
>>41691252can you please post the screenshot or whole text of this error?
>>41691267 >Am I getting filtered? I don’t think so. At worst, try using an empty preset to check. It seems more like a proxy or configuration issue. Can you screenshot how you’re connecting to Gemini? (Hide your key if you’re using a proxy.)
>>41691302So I downloaded a gemini-specific JB and it finally pushed through a SFW convo, but blanks return the moment I try to edit anything in the JB. Definitely feels like I am getting filtered to shit. Maybe I'm missing something?
>>41691396 First, if you have access to MM, you have access to the very latest Gemini model, which is often preferred. https://files.catbox.moe/1g1pdo.PNG If you don’t have it, you can either: - Take the latest Tavern staging version - Add this inside public/index.html: `<option value="gemini-exp-1121">Gemini Experimental 2024-11-21</option>` https://files.catbox.moe/xfxzqm.PNG - Or use it like this with external models in your current version: https://files.catbox.moe/ctpu5a.PNGTurn streaming off if you have issues with repeated lines at the end. For presets, there is: https://rentry.org/gemini-qr#presets I didn’t test Gemini that much, so I can’t recommend which to use. You shouldn’t be filtered unless you’re doing cunny, but even then, I heard it can be bypassed. I also recall discussions saying that disabling streaming could help. From your screenshot, the only thing I’m unsure about is the Squash + System Prompt enabled.
>>41691450Yeah, I don't have 2024-11-21 in options. Guess it's time to update again.
>>41691457You can do the third options if you don't want to update. I had some filtering problem using google AI studio as a source too, but none while using OpenAI with external models.
Man I love posting random shit for an AI to make sense of, it's like the digital equivalent of kicking a robot dog, watching it right itself and then kicking it again.
>>41691529Spitefag is that you?
How does your favorite pony react when you suddenly drop everything you're doing and sing the theme song?
>>41691632
>>41691632And second.
>>41691786Model / JB?
>>41691795Model is on the log, it's Chorbo.Preset is here >>41663387
>>41691810Thx
>>41690559 Added, thanks anon. https://mlpchag.neocities.org/view?card=littlewitcher/Marezilla.png The website is apparently now at more than 500,000 visits, which I suppose must mostly be bots because I don’t think it’s known outside of here.
>>41691961Sad
>>41692050 >More songs I want to sing them songs like "Gypsy Bard" or maybe "This Day Aria" to Cadance now.
>>41692095>>41691632Its kinda cool that it knew that one. What's the most obscure pony song you think it knows?
>>41692050>>41692149which model can vomit out this much soul
>>41692165 NTA, but from the log name, it’s Chorbo/Latte/Chat-GPT4o-latest. >>41692149 Really neat that it knows Gypsy Bard.
>C.O.O.L
>>41692149The RDP sinking ships songFamous in the fandom(ish), but not really known as widely
Is it worth paying for unlimited Claude 3.5 haiku but it's unlimited RP?https://desuarchive.org/mlp/thread/41621795/#41628832
>>41691786sunbutt a cute
>>41692623No, we have better local models at this point and will likely have far better ones that than in the next year.
>>41692623If it was Sorbet (or hell, even 2.1), I would consider it. But Haiku isn't worth it.
>>41692652You do realize that it's lifetime right? So when Haiku is sunsetted for a better corpo Claude model it will only get better. Local models can't keep pace forever with even the shittiest Corp api
>>41688419>has really good context and able to have tell a coherent story on 30000t+,Doesn't it have like 32k max context or is that MM only?
>>41692748or local start to get trained by us like they did with the pony model for image generation and it gets more apeal and open source training than corporate and becomes more relevant than the censored ones that keep getting kneecapped every time an alarm ring in their department, because its not they don't know you guyz are fucking horses on their machines, they know, but like a ban wave they just wait up to clean up everything and cut the lifeline to thousands at once so that people dont slowly adapt to the new changes and everyone is at a lost at to what cause the bans and censoring since you cant test with new accounts to see what go through and what doesn't so that the model becomes the perfect little spreader of "the message". i still believe community made models will always surpass corporate ones when they work together to give everyone what they want rather than what the corporate want us to have
>>41692748>Things that seems good right now might be irrelevant in 6 months, its always bad to invest long term considering how fast technology changes nowadays and its best to be able to pull the plug before you lose too much or get stuck holding the bag because you've made a long term agreement on something that will be outclassed by something we cant predict. thats why they do these lifetime deals, they know most of the customers wont even use most of the amount the initial cost was supposed to cover, or they might even turn around and say, lmao just kidding, server upgrade require more money so we cant hold our side of the bargain and need to charge you again.
>>41692907SOURCE : any triple A game with a roadmap of release, dont make enough money, then say lmao, and just shut down the game for the people who paid
>>41677636https://files.catbox.moe/cxkcph.pngPonyville Cooking ShowDashie and Flutts have a cooking show where they make pony food and invite guests onWorks best with a JB that does the "pull more info about character" instruction probablynot mine card. another anons asked me to drop it there
>>41692920is the lore book embedded in it?
I want a cute reformed changeling card.
>>41693003fren changeling, basically a horse ditto on your side
>>41690559>>>/trash/
>>41692907AI companies make money only through constant usage and Anthropic likes having us around and their employees are part of the roleplaying community. Your shitty local model cannot compare to the $4B that just got invested in Anthropic AGAIN this month. US corporate AI models are the best in thr world and are advancing faster and faster because they now see it as an arms race with China. Your homebrew is the equivalent of thinking it's easier to get a ride to space by making your own space program versus hitching a ride with SpaceX/NASA/Roscosmos Soyuz and benefitting from their advancements.
>>41693048Stop namefagging when it’s not relevant, please.
>>41693094He's new, go easy on him
>>41693103 I said please, and I didn’t call him a fag. That’s the equivalent of a friendly tap on the back, really.
Remember 20 threads ago when we were trying to make our own AI startup like /aicg/ did? Good times. I still chuckle when thinking about having that German freelance AI engineer sort through horse shut to build training data
>>41693130I had a rich friend interested in providing VC funding
>>41693130 >We were trying to make our own AI startup like /aicg/ did? I believe it was mostly wishful LARPing. But the image of the German freelancer was indeed funny.
What's the request model name for Sonnet 3.5? Risu changed their website setup
Time sensitive email sent
>>41693330Email? Are you talking to someone here?
>>41693092>likes having us aroundYeah, that's why they put in a defense that hasn't had Claude on any proxies for 2 months now.If it suddenly seems like the corporations are on your side, it just means they don't know enough.
>>41693286claude-3-5-sonnet-20241022 or claude-3-5-sonnet-20240620.https://docs.anthropic.com/en/docs/about-claude/models
>>41693330Thanks for the heads up!
>>41693346They like having us around. AWS doesn't because they don't want their customers to leave. Amazon instituted the security measures because that's all on their side. Anthropic had been getting notifications for months about this but didn't tell Amazon and gave us more time
>>41693330Replied. Sorry for the typos and the awkward wording, it's pretty late over here.
>>41693330>$30 monthlyThanks but fuck off.
claude-3-opus-latest
>>41693492Why are you blaming him? It's jp setting the price
>>41693648Is that a request for her?
>>41693092Delusional post.
>>41691489That worked, thanks. Gemini is pretty damn good. Maybe it's just Claude withdrawal speaking, but after latte it's like a breath of fresh air.
>>41693668Who even is she?
>>41693492>not racking up hundreds of dollars worth of chats every dayDYEP?
>>41693846Alt universe Chrysalis from the Reflections arc of the IDW comics. No more than a background cameo in the actual story, though.
>>41692894I WISH TO KNOW
>>41693886Size 2
>>416938861114 and 1121 'only' have 32K. Yeah.
>>41692894Gemini "by default" has 2m context on Pro lineup of models. Gemini 1114 and 1121 have only 32,000 context but it will be expanded later to 2m as wellthose 2m tokens are not real-real, they are mostly to process audio and video. for text-based generation you are still stuck with ~35.000 effective contextmy line was more about that Google holding its context better than Claude and GPT and able to recover facts from it more effectively. even on 30,000+ chats I saw Google recalling events from 20,000+ ago and been able to use them creatively in text. something both Claude and GPT are no able to do on its own>>41692925no, no lorebook
>>41694036>those 2m tokens are not real-real, they are mostly to process audio and video. for text-based generation you are still stuck with ~35.000 effective context2 mil is for long document processing as well, even if the performance is reduced it's still better than nothingGemini is trained on the examples of at least 128k length, and holds up until at least 128k, it's generally way better at handling long contexts than others (see RULER) and has way less degradation in the middleyou still need summarization despite all this though, as it tends to forget complex things even if it recalls simple ones
>Wrote, I think, ideal mare for me.>She's scandalized and sarcastic towards me.
>Column still dead
>>41694241 Maybe that’s what makes her the ideal mare.
>>41694294It was alive last night, ~12 hours ago. Source 3 is kind of wonky and sometimes cuts off in the middle of generation, but its was doing alright.
>>41693705he's not delusional, he actually correct in some aspect, every post someone make just feeds the machine and they get to make free data out of it and people are willing to pay to be the product just so that months from now they change the formula on em. remember patreon or netflix? make it big by seeming consumer friendly, then shift and lock people in a shitty ecosystem.
>>41689578got some free time to try a bunch of new models, i have mixed feelings about qwq, yes it is indeed next level smart and yaps about lots of good knowledge of ponies and animal crossing and other things, but god that thing cant fucking keep to itself, it just has to talk for {{user}}, i put rules against it in the greeting, in the char card, in the system prompt and the instructs and it still fucking insist on speaking or acting for {{user}}, for me that's a deal breaker.
>>41691632I was about to>rapemy filly self.But it seems it was consensual after all.What a faggot he (now she) is.
>>41693330Thanks for keeping us updated. And I hope things will get better for you soon, hang in there, friend! >>41693492Well, I agree that paid proxies are pure cancer (links to voluntary donations do not count). But just being aware of what's happening around is still useful. At least we now know that some keys can still be found and used, since j has some. (｡︵｡)
>>41693619Entire purpose was to have a proxy for ourselves not a proxy to paypig for. Simp
>>41694723 Cute art, as always.
>>41692920Added, thanks to the anon.https://mlpchag.neocities.org/view?card=Anonymous/Ponyville%20Cooking%20Show.png
Where is a G4 lorebook?
>>41695201 There isn’t one because most models know enough about G4, so it’s not really that necessary. A lorebook, if made, would be more about small details they can still get wrong, mostly from the later seasons. But even those, newer models are better at handling. I remember when they couldn’t tell who Autumn Blaze was and would hallucinate a pony themed around the season. Now, if you ask Chorbo, it gets it right. They could get the race of background ponies wrong too, but now I don't see it happens.So if you want to make one, you first have to consider: - Which models would I use it with? - What are the things they still don’t know or get wrong?
>>41695286>A lorebook, if made, would be more about small details they can still get wrongLorebook for horsefucking and body parts?Or maybe, a wider vocabulary converter.
>>41695286Oh. Well that makes sense.
>>41695201G4fags are just happy with the same slop
>>41695338 >Lorebook for horsefucking and body parts? I believe most people include a section for that in their preset. If you put it in a lorebook, you’d also need to decide when it activates. You could set it to always be on, but then you might as well use a preset’s section. Alternatively, you could try to activate it with specific trigger words like "sex," "marehood," etc., but that might be clunky. >Maybe, a wider vocabulary converter. Same problem as above, you’d have to either always have it on, unless I’m missing something that could work better.
>>41695380QwQ already knows equine anatomy really well. If you even ask it, it goes into detail on teats being under the belly and how equines would "do it"
>>41695380Well, I was thinking, it makes more sense for world-related stuff to be in lorebooks, like everything mlp.Then, you could use the character tags to activate the lorebook.So, the lorebook acts as a preset and is always on on ponies but it's auto off on other characters but can still trigger if you start horsefucking a non-mlp character.
>tfw dancing with Luna and Celestia
>>41688833One day left to go if you missed it.
I see a lot of comments about QwQ.How does it compare to the online stuff?Does it put a good fight against gemini 1121?
>>41695774I would say its smarter than anything not 3.5 sonnet, it obviously knows less about the fandom but its smart enough to use what you tell it intelligently instead of reusing it word for word.
>>41677636Made an EqG Applejack bot, with five scenarios!>You run into Applejack at the school bake sale>You meet with Applejack bright and early to work on helping repair her family's barn>Applejack approaches you after class to ask for tutoring in physics>You run into Applejack in downtown Canterlot, where she's working part time handing out fliers for Flim and Flam's store in a banana costume>You and Applejack sneak into the barn at night to mess aroundhttps://venus.chub.ai/characters/Maverick074/applejack-380a9ecdd7beI also updated my EqG Rarity bot to add an additional scenario:>Rarity invites you over to her home to model swimsuits she's designedhttps://venus.chub.ai/characters/Maverick074/rarity-ea28057b2b49
Gonna recommend this QwQ merge for anyone doing local.https://huggingface.co/bartowski/EVA-QwQ-32B-Preview-GGUF
Is there a difference between Chary and MM 4o?I have access to both but just wondering if there's any downside to using one over the other?
>>41696372Maybe the max context size? But I don’t know what MM is using for it.
Is it possible to chat with my wife on an iphone with ST or an equivalent app? Im assuming probably not, but figured I’d ask.
>>41696571Yes!
>Typ-delay/watch-workpossibleCASEQUENCETOFOLLOW<Sorry!] ( !resume.)>GPT spilled the beans of me being watched/logged by the typ-delay, probably for assisted learning in neural network>Or maybe I am just schizoProbably the latter, but it certainly felt more...free, than usual. I'll just stop for the night I think
>>41696575Oh really? How so? I see there’s a guide for Android in the OP, but that doesn’t help me.
>>41693330Sent out follow-up>>41694723I suppose I just have to take things one day at a time but all the well-wishes mean a lot to me, thank you. Also, cute!>>41693357You're welcome!
>>41696618 NTA, but for iPhone, I suppose you’ll need to use something like Risu (In the OP). The only problem is that you’ll have less support since fewer anons use it, and all presets/JB files are made for Tavern. So you’ll need to create your own or port them somehow.
>>41696099What happens if you put her and normal AJ in a GC?
>>41696673Risu community grows stronger every day actually
>>41696686Try it and post logs
>>41696324with only 12GB of VRAM, is it even worth it to try local?
>>41696879I think 12gb is sufficient for decent models
>>41696883but according to an LLM VRAM calculatorI can only run that at "IQ2_M" or less, whatever that meansif I could use my system RAM too, I'd be able to run most of them, but won't that heavily impact the performance?
Fuck fuck fuck fuck fuckhttps://techcrunch.com/2024/12/01/aws-launches-an-incident-response-service-to-combat-cybersecurity-threats/
>>41696958 Claude on AWS was already fucked. From reading the article, I don’t think the tool they’re proposing worsens the current situation much.
Gepetto.>>>/g/103377853
>>41696894It's not a terrible performance hit if you only use some RAM. 22B models should be very usable at Q3, especially at lower max context.While I haven't optimized for 12 VRAM, if I were you I'd try Nemo since that can entirely fit on your GPU, probably at Q5. There's a pony finetune, as well as stuff like UnslopNemo.
>>41696958Nothing lasts forever, I guess. Looks like it's time to say goodbye to pr*xy Claude for good.>>41694801<3
Blank responses on Column again. The fix on the rentry doesn't work either. I get the feeling it's showing source 3 but it actually is source 2
I have reached the maximum context allowed in one of my chats and now that I'm trying to learn more advanced stuff I wanted to ask, what should I do? If I delete the older prompts and clear chat history, won't it cause the bot to act weirder or go off the rails without the previous information?
>>41697190The solution is simple. Have it summarize your chat in an itemized list (or summarize it yourself) and make a new one, paste the summary above, and continue from where you left off.
>>41697190>>41697212I did this once. It worked pretty well. The quality rose a bit for a while.But my chat wasn't all that intricate, this runs the risk of losing details.
>>41697220I'd recommend using something with a high context limit like chorbo, but yeah, even that can lose details. Hence why I suggested the option of also summarizing it yourself. If you like your chat that much, should be no trouble to just look through it and takes notes, then write up your summary and put it in the next chat.
>>41697212>>41697220Doesn't ST already come with a function to summarize? could that work?
>>41697223Also I forgot to mention, I'm using Chorbo, but the proxy I found has a 32000 max context limit, that's why I'm asking.If not for that limit I hsould be able to continue my roleplay up to infinity, right?
>>41696879thats what i run it on, you'll need a bunch of actual ram to load the model on tho>>41696324idont know how to make qwq stop fucking talking or acting for anon, i dont like that model for that one reasonalso i tested a bunch of other models and will be updated my rentry later today, found 3 new awesome models to share
in the meantime before i update the rentry, anyone interested in the models, check the 4 at the top, its the ones i uses now on 32 gb of ram and 12 gb of vram
an example on the pony i made and how well it respond
>>41697266Many thanks on your experiments, I gotta check those models out.How do you run these local models?I also have 12 GB of VRAM, but I'm starting out with LM Studio and I can't figure out if it's using my VRAM or RAM and how to mix them.
>>41697275>How do you run these local models?There is a guide in the OP.
>>41697275>How do you run these local models?i made an entire rentry post, the link is in the OP, explaining how to run it locally while simplifying the tutorial as much as i could, i will overhaul the tutoriel and model documentation when im back from work tonight but koboldcpp_cu12.exe allow you to run a model locally and split the load between VRAM and CPU, the normal ram is just to hold on the model and context limit, i usually set mine at around 32 000 context token
the following models were added to Chary proxy:- Gemini EXP 1114 32k context (max)- Gemini EXP 1121 32k context (max)- GPT 4o 1120 64k contextGPT 1120 in my tests filters less than Latte but YMMV. among two new Gemini - I don't see much differences between them so test them both. and if you are familiar with previous Gemini - then YES two Gemini are slower, it is not a proxy issuecorrect names for index.html:>gemini-exp-1114>gemini-exp-1121>gpt-4o-2024-11-20if you don't know how to edit index.html then download new ST 1.12.8 which is already compatible with those models
>>41697295>Gemini EXP 1121 32k context (max)Nice.
all the models tested so farWhite : i keep, had a choice to make due to hard drive spaceblue : usable, good enough but i didnt keep themyellow : usable but had visible flawsred : unusable or massive flawspink : unusually horny
>>41697250Not really. Models tend to have a soft context limit and a hard one. The hard cap of any model is the absolute maximum that it can process, but their understanding of history tends to suffer a sharp quality loss after a precise point far before that. This happens because their effective limit is this lower number, and anything over that is just some clever tricks. 32k is about right for most current models (and I think Chorbo as well); you shouldn't go over that. That's why proxies usually limit you to the effective context anyway—no point in burning 4x the tokens on something with little to no gain. But to answer your question: IIRC the max that Chorbo can handle is 128k, though you really should limit it to 32k.
Chorbo ponified 'Albert Einstein' into 'Albert Equestrian'. found it vaguely amusing
>>41697399>not albuck einsteed
>>41697399Old log but I liked Sigmane Freud too.
>>41688833Poll closes in 2 hours.
>>41697430Hot damn, how did you get a log that good?
>>41697439 It was back in the Claude Opus days. I don’t remember the preset, but it must have been a modification of the Pratchett/Vonnegut one.
I have a strange feeling that all the Gemini exp models are getting a little dumber(also more filtered) with each passing day. :/
>>41697509 I don’t think they changed the snapshots, but I had some header messages yesterday on SillyTavern with MM telling me it couldn’t generate due to "Reason#1" or something like that, and today no problems. So maybe it’s something with the keys? If you have another proxy with Gemini, maybe try it. Gemini 2 is expected to release this month also, so I hope they improve even more.
>>41688833Here are the results. A tie and generally quite close outcomes. So I'm making another quick poll to find out which theme you want first, knowing that it will be Theme 1 -> Christmas/New Year -> Theme 2.https://strawpoll.ai/poll/vote/UIyOK7cPxLFJ
>>41677636A lazy card based on the >>41682841https://litter.catbox.moe/3g34hh.png
>Ask Gemini how he would describe a mare's pussy.>"Imagine the velvety muzzle of a foal.”Not sure that's the best image, but you do you, Gemini.
>>41697790Really hoping it's locations first, because I've got some cards in my back pocket that match the theme!
Someone asked about it before but an issue issue I'm having is the AI gets stuck wanting to go over the last 2-3 messages instead of the most current one. The way to fix it was something to do with XML wrappers. Looking at it I believe the Knotty JB has something like this already or am I just reading it wrong?
>>41698113 I’ve never experienced this, but from what I’ve heard, it can happen with heavy presets like Knotty. The post you’re referring to is this one, I believe: https://desuarchive.org/mlp/thread/41638289/#41657036 What you can also try is this: https://desuarchive.org/mlp/thread/41568022/#41588543
>>41697797>Blacked cuck shit
>>41697295>download new ST 1.12.8Is this safe? Has the discord tranny calmed down over nuking proxies and rp?
>>41698176kek. didn't check myself (I am still on 1.11.6): but a) haven't heard any complains, b) don't see anything suspicious in changelogyou can download a separate ZIP folderhttps://github.com/SillyTavern/SillyTavern/releases/tag/1.12.8click on Source code (zip)unpack it somewhere else, launch, check the features, if things are fine, then update your main ST installation to 1.12.8 as well
>>41698176 He either took his meds or accepted that something called 'SillyTavern', a fork of something made especially for RP, is, in fact, associated with RP. I downloaded 1.12.8 and see nothing made against proxy management or RP. Nothing in the changelogs either.
Is MM GPT4 still working? I tried pasted the link but I keep getting errors. Does /proxy not work as an endpoint anymore?
>>41698195proxy/openai for all GPT modelsproxy/google-ai for all Gemini models
>>41698195Ope, nvm I figured it out
>>41698195Getting errors on Gemini. I think the server might be fucked, I'm not even getting a connection, though the cloudflare link still says its up when visited.
>>41697326I see, thanks for your answer. So creating a new chat with the same character and then pasting a summary of the previous one is the only option? or are there more?
>>41698210Dankie anyways
>>41698213Gpt works
>>41698155>it can happen with heavy presetsBeing big and better has its downsides. I guess I'll try out the couple of experimental options and those listed from the other thread.
>>41698214You could try using the summary tool to keep some stuff in tact. here's a prompt i pick up awhile back. Just remember hit summarize every now and again. Or set it to automatically to do it every chunk of messages.[Pause the roleplay. Right now, you are the Game Master, an entity in charge of the roleplay that develops the story and helps {{user}} keep track of roleplay events and states.] [Your goal is to write a detailed report of the roleplay so far to help keep things focused and consistent. You must deep analyze the entire chat history, world info, characters, and character interactions, and then use this information to write the summary. This is a place for you to plan, avoid continuing the roleplay. Use markdown.] [Your summary must consist of the following categories]: [Main Characters]: An extensive series of notes related to each major character. A major character must have directly interacted with {{user}} and have potential for development or mentioning in further story in some notable way. When describing characters, you must list their names, descriptions, any events that happened to them in the past. List how long they have known {{user}}. [Events]: A list of major and minor events and interactions between characters that have occurred in the story so far. Major events must have played an important role in the story. Minor events must either have potential for development or being mentioned in further story. [Locations]: Any locations visited by {{user}} or otherwise mentioned during the story. When describing a location, provide its name, general appearance, and what it has to do with {{user}}. [Objects]: Notable objects that play an important role in the story or have potential for development or mentioning in further story in some big way. When describing an object, state its name, what it does, and provide a general description. [Minor Characters]: Characters that do not play or have not yet played any major roles in the story and can be relegated to the 'background cast'. [Lore]: Any other pieces of information regarding the world that might be of some importance to the story or roleplay.
>>41698214There's no reason (other than saving tokens, but that's only relevant for non-proxy usage) that I can see to start a new one.You could put the summary into the author's note section and continue in the same chat. Or use the auto-summary tool, but I've had mixed experiences with it, back when we were all coping with 8k context on models.Silly Tavern is pretty smart and, if you set the context limit to 32k, will automatically cut some the oldest messages to fit inside that limit. So you'll only gradually lose the non-summarized details, rather than losing them all at once with a new chat. This is generally better.
Seems relevant and something I had on my mind. Is there any negative to having the summary set to automatic all the time if using a proxy?
>>41698024Weird foalcon word choices aside... wew.
>>41697797>>41698164It didn't catch my attention at first, only really caring after anon shitposted, looking to see how bad the blacked meant. Seems like a simpler estrus sim but with a focus of ponies having taken over south usa (Why you didn't change for earth I have no idea).Still, makes for an interesting change in comparison, as the other has you in Equestria, though it's also better crafted if we may be fair to both cards, but editing for earth, you can place your house in a pletora of places to start out to make for an interesting hide away or trying for escapades. I wonder how the card would try and write zone 51, but currently trying a south american jungle and harvesting berries and shit.
>>41698263Not really...besides it constantly rewriting its memory like an dementia patient, I guess it should be fine?
>>41698263>if using a proxyproxy has a ratelimit, say 4 messages per minute. your summary request (automatic) counts as a separate message. so you will unavoidably hurt your own ratelimit this way>why does it say out of quota? I sent only two messages!>because two other messages were summary that you never read, never bother to check what ST even writes into it, and unsure whether model even gets right
>>41696099Added and updated.https://mlpchag.neocities.org/view?card=Maverick074/Rarity.pnghttps://mlpchag.neocities.org/view?card=Maverick074/Applejack.pngThanks anon.>>41697797Added too. Thanks.https://mlpchag.neocities.org/view?card=MaudPie/Earth%202045.png
>played with vpn>cockblocked due to IP limit on columnI feel dumb
>>41698263summary doesn't work well on auto because it runs in random points of the story and doesn't know how your current unfinished scene endscheck https://rentry.org/how2claude for the summarization advice, helped me a lot in making my summaries more coherent
>>41698390Send a mail, I'm sure he will understand.
Just noticed that one ai site that was been shilled here was a rugpull. They copied it and made another one slightly different as well.rubii.ai
>>41698470 The best approach is just to ignore it and never trust websites like this.
Is column still alive? If so, does it have aws?
>>41698516>does it have aws?No. We are back to the days where we leached from random services.
>>41698541Do you know how long aws lasted the last time on it?
>>41698552It never used aws.
Oh...
>>41698490Tavi, my beloved...
Updated the rentry tutorial for local hosting and models i've tested and updated new setting file for formatting and context and new default model.>ST-formatting.jsonI've also added what i've learned here about how the model works and how ram / Vram relates to the models and what you can run.Feel free to check out the new model experiments and new formatting master import file.https://rentry.org/lunarmodelexperiments
Hey guys, my messages are being removed from context too quickly and the bot is only reading ~7 chats and throwing out the rest.It's not token limit because I've trimmed a few thousand tokens and it's still dropping.Where is this retarded setting to change this.
>>41698703
>>41698703there's 3 possibility, but first are you local or proxy? if proxy, i cant help you, never tried itif localcheck the launcher, its possible you don't have the token input limit loaded in the model loaderthen check in the slider tab at the top for text completion if the token context is also the same as the ones from the loaderand lastly, swap model see if the problem persist, its possible that the model is just dumb when it comes to memory and have a limit on itself
NEW BREAD>>41698699>>41698699>>41698699
>>41698718>>41698720It's not the context it's going message by message ignoring how many tokens they take upIt is only reading the last 7 messages
>>41698746 Then I don’t know, man. I’m not even sure Tavern has an option to do that. Did you try changing preset to see if the problem is here or not.If it’s truly fucked up and you don’t get an answer from another anon, just download a new SillyTavern and move the `data\default-user` folder from your old folder to the new one. That should keep your bots, chats, and presets.
>>41698795I'm getting 8 messages now after adding every single lorebook I have savedI have no idea what's happeningI've been avoiding updating ST after there were rumors of them disabling proxy supportI guess I can try updating it before I go and download a fresh one
>>41698807Make sure to back up your bots, chats, personas, etc, first.
>>41698470What was the original?
>>41698256>There's no reason (other than saving tokens, but that's only relevant for non-proxy usage) that I can see to start a new one.The longer you go on for, the worse and more repetitive the AI becomes. Starting a new chat resets the repetitiveness of its writing and keeps things fresh.
