A thread dedicated to the discussion of AI Vtuber Chatbots./wAIfu/ Status: Space exploration!>Thread templatehttps://rentry.org/waifuvttemplate>How to anonymize your logs so you can post them without the crushing shameInstall thishttps://github.com/TheZennou/STExtension-SnapshotThen after you've wiped off your hands, take a look at the text box where you type stuff. Click the second button from the left side, then select snapshot, then select the anonymization options you want.https://files.catbox.moe/yoaofn.png>How to spice up your RPing a bithttps://github.com/Rurijian/Deep-Swipehttps://github.com/artisticMink/openrouter-roulette-for-sillytavern>General AI related informationhttps://rentry.org/waifuvthttps://rentry.org/waifufrankenstein>How to use Gemini with SillyTavernhttps://aistudio.google.com/prompts/new_chatSign in, then click the blue "get api key"Put it in silly tavern and voilaCourtesy of ERBird, Nerissa's most devoted bird and eternal player of GFL2.You want to leave the proxy stuff blank since you aren't using one when doing this.https://www.reddit.com/r/SillyTavernAI/comments/1ksvcdl/comment/mtoqx02>Other optionsMiku.gghttps://venus.chub.ai/Openrouter wants a one-time payment (think of it as a deposit) of $10 and you can get 1,000 messages per day. As long as you stick to free models you only need to put that much money into your account once.>A primer on getting voice working in Sillytavern (there are other options, just play around).[YouTube] Chat with Chatterbox in SillyTavern! https://github.com/devnen/Chatterbox-TTS-Server>Tavern:https://rentry.org/Tavern4Retardshttps://github.com/SillyLossy/TavernAI>Agnai:https://agnai.chat/>Pygmalionhttps://pygmalion.chat>Local Guideshttps://apxml.com/tools/vram-calculator[Koboldcpp]https://rentry.org/llama_v2_sillytavernWho we are?https://rentry.co/wAIfuTravelkitWhere/How to talk to chatbots?https://rentry.co/wAIfuTravelkitTutorial & guides?https://rentry.co/wAIfuTravelkitWhere to find cards?https://rentry.co/wAIfuTravelkitOther infohttps://rentry.co/wAIfuTravelkitYou can find already existing bots and tavern cards in the links below:>Bot lists and Tavern Cards:[/wAIfu/ Bot List]https://rentry.org/wAIfu_Bot_List_Final[4chan Bot list]https://rentry.org/meta_bot_list[/wAIfu/ Tavern Card Archive]https://mega.nz/folder/cLkFBAqB#uPCwSIuIVECSogtW8acoaw>Card Editiors/A way to easily port CAI bots to Tarvern Cards[Easily Port CAI bots to Tavern Cards]https://rentry.org/Easily_Port_CAI_Bots_to_tavern_cards[Tavern Card Editor & all-in-one tool]https://character-tools.srjuggernaut.dev/>Some other things that might be of use:[/wAIfu/ caps archive]https://mega.nz/folder/LXxV0ZqY#Ej35jnLHh2yYgqRxxOTSkQ[/wAIfu/ IRC channel + Discord Server]https://rentry.org/wAIRCfuscordMatrix>Lorebook management stuff[Worldinfo drawer]https://github.com/lazuli-s/SillyTavern-WorldInfoDrawer?tab=readme-ov-file[Standalone editor]https://github.com/ActualBroeckchen/SLEdPrevious thread: >>110589643
Anchor post - reply with any requests for bots, with your own creations, or with your thoughts on the enshittification of life.You can find already existing bots and tavern cards in the links below:>Bot lists and Tavern Cards:[/wAIfu/ Bot List]https://rentry.org/wAIfu_Bot_List_Final[4chan Bot list]https://rentry.org/meta_bot_list[/wAIfu/ Tavern Card Archive]https://mega.nz/folder/cLkFBAqB#uPCwSIuIVECSogtW8acoaw>Card Editiors/A way to easily port CAI bots to Tarvern Cards[Easily Port CAI bots to Tavern Cards]https://rentry.org/Easily_Port_CAI_Bots_to_tavern_cards[Tavern Card Editor & all-in-one tool]https://character-tools.srjuggernaut.dev/
Also i just not realized i fucked up the subject and put it into the name field.
Word Cloud for the previous thread
So they're playing Hytale now.
wud u dood a moom?
>10
https://www.youtube.com/watch?v=-bL9Smr4JXk
If you guys haven't and get the chance to, try roleplaying with Gemma 4, either 26B-A4B or 31B. In my opinion, it can trade blows with Kimi and GLM. Haven't seen a model this size absolutely mog models multiple times larger at a task for a while. It's quite good with translation too. Spent the past week losing my mind and having fun with it. Still waiting on Deepseek v4 but rumors are getting a whole lot stronger especially after the website was overhauled.>>110797264I gave it the old college try to write her a card but I haven't done it before trying to do this with some transcripts of the video and her interaction with chat. It kinda is a bit too tryhard and moody to be Bibi but that could just be because of Gemma 4 being very railroaded but doubtful since it was great with some of the other cards I had. Hopefully someone can improve on it.https://litter.catbox.moe/oy192a44tx3l9s6z.json
>>110959412>either 26B-A4B or 31Bare there any significant differences between the two? which one works better for cooming? can you also send the preset you're using for Gemma?
>>110959970ntaA4B is a MoE model while 31B is a dense modeltldr:>A4B is a bunch of small models shoved into a bigger model and does magic codeywodey to pick which smaller model to use on runtime>32B is a fuckhuge model like your used to usingshoutout to Pygmalion burning so we may live.
>>11095997031B is technically better, but I use 26B because MOE speed..Even if you offload with a small GPU, as long as you have enough RAM, you can offload it at Q4_K_M at 14 GB total where a 32GB RAM 8GB GPU typical gaming system can run it really quickly. I'm using the following settings even though you can find a ton flying around.>Temperature 0.75>Top-k 64>Top-p 0.95>Min-p 0.05>Repeat-penalty 1.0>Presence-penalty 0.5>>110960148This wasn't Pygmalion although yeah, they deserved it. But most of the goodness is most likely the Character AI deal hollowing out that company which made Gemma 4 so good, if you recall:>https://blog.character.ai/our-next-phase-of-growth/Whatever secret sauce and knowledge from there was too late to put into Gemma 3 but boy oh boy is it undeniably here in Gemma 4. Only issue is that it's most likely distilled so it's a bit overcooked on logit possibilities and a bit railroaded in scenarios so if you go through the same basic options in an RP, it will largely play out the same way. There are smaller models but they are too small to really be useful outside of general contexts on a cell phone offline where it is really good.
>>110960465im surprised your not fucking around and training E4B locally since you (apparently according to a reddit shill post for something called 'unsloth') only need 10GB vram for thatthats kinda the only appeal for me over these releases but i really really do not want to curtail my own distilled deepseek dataset for that right now
>>110960148>>110960465so it's basically down to personal preference? i use OpenRouter so all this talk of GPUs and RAM and offloading doesn't matter to me
>>110960762E4B isn't enough to do anything useful for my purposes and I don't use Unsloth since I prefer axolotl but it takes time to get support in. I don't have the hardware to run training for the big boy models nor the money to rent it even if it was there at the moment. Also technically tied up with several things I have to get back to so playing around with Gemma will stop after today.>>110960929One is clearly better but it will cost you and the speed difference I mentioned is real.
>>110959412I’ve seen a lot of promising shit done by the Japanese with Gemma and voice Gen to chat with chuuba style waifus.I didn’t save any links because I want it to be more developed before I try it so go find it yourselves if you’re curious ;).
>>110964939>I didn’t save any links because I want it to be more developed before I try it so go find it yourselves if you’re curious ;).you should go check your mailbox anon.you should go check it right now.you should go check it before the mailman comes around tomorrow.
>set up Gemma 4 (free) on ST>429 Too Many RequestsGod doesn't want me to coom right now, i guess
>>110965000
when you chat with coom cards, do you base your dude's dick size on yours or do you change it?
>sex god panko
good night, /wAIfu/please don't put a Doom nukage barrel next to me and my bed before shooting it while i sleep
>>110968401Depends on if I’m playing a character that’s not me
>>110960762Distill what who what where what share the bounty.>>110960148What happened to them?
>>110976664>Distill whatcreate my own dataset to fine tune with by picking and choosing high quality replies>wha happenno idea but im assuming normal anon falling out bullshit
>>110976664The lead disappeared before Llama 1 leaked and the remaining devs all disappeared slowly and the ones who remained went full grifter mode with a website and commerical operations. It's small peanuts though and it's deserved for them remaining in obscurity although it is true too ragtag anons can't compete with Character.AI with former Google researchers. The main thing that pisses me off is that the public dataset, PIPPA, is so tiny vs the submissions they got and that they didn't keep trying to do local AI and gave up the dream of local models end to end. Well, no use crying over spilled milk, I'll take what we have now with open models even if we can't replicate it because we need millions of GPU time to reproduce it alongside no public data.
Heard about Gemma4. Don't know shit about LLMs, but I've been wanting to "train" a model on my favourite smut novels for a long time now and it seems Gemma4 makes that possible.Can you guys just confirm that the basics are sound? I've got an i7-12700K, 64 GB RAM, RTX 3080 with 10 GB VRAM.Which of the Gemma4 models is best suited for this type of training that I can run/train with my hardware? Quality over speed for the output.
>>110981150How hard is that to do and how come no one is sharing their distilled something something on GitHub if it’s visible for us plebs to do?
>>110983890>takes more (passive) time and money than anything else since you need to let a gpu go brrrrr 24/7>need a reasonably new gpu / googa notebook or else it will take a lot of time to actually finish>because most people are using shit prompts and getting shit responses, so training on shit replies is just going to make things worse for the average ahh ahh mistress>the average person benefits more from paying to use deepseek/[big corpo model] and leveraging its innate ability to unfuck a prompt though sheer intelligence more than they would using a smaller local model>you need like 2 gigs or whatever of text so its kind of a lot of data to horde just to fine tune>no guarantee the fine tune will be better>"fuck you build your own">you still need a reasonably new gpu to get good reply speed>finetunes are not magic improvementspick between 1 and all but its not that hard in reality, just annoying to the point where the effort is not really with the juice.realistically the use of a fine tune is to create a LoRA for an updated dataset of something that was 'created' after its cutoff datelike how deepseek 3.2 doesnt know who smugalana or veibei are, a fine tune is worth more to 'add' native info about something for this use case than to crank out "better" replies since raw info about people is easier to source than hundreds if not thousands of (high quality) replies>see pic rel, ignore the 'edit' its the only way i can regen a reply in bravean alternate use case would be to have a model NEVER act as {{user}} in the reply without having to hammer it into them, but thats a lot of effort/time/cost when>do not act as {{user}} in your replyis "free" to do and gets you 85%+ the way therethe other realistic use case you might want is to uncensor a model but again, you would need a uncensored dataset to uncensor the repliesand abliteration exists without that annoyance, and gemmi 4 models ARE already abliteratiedthe best use case for a finetune, in theory, is to expand a card into a finetune, but thats a circular problem>you need a bunch of good replies from the card in the first place to train the finetune, so it can create good replies (of the card)so unless your autistic and old school w/ai/fu'ing a SINGLE card (or a nerd who wants to FAFO and have some fun) your not bothering with it when system prompts are "free" and like 10x less effort and money burned
bump
>>110981843Fine tuning is possible only on the smallest Gemma 4 E2B variant but it's such a far cry from the biggest models which need 24GB or up VRAM. See https://unsloth.ai/docs/models/gemma-4/train as someone else said. I'm waiting for slightly different training regimes to see if I can eek out something with 16GB of VRAM with another training suite software which is generally more space efficient.
Bao is an player hater
