I have tried candy.ai and it is very good but after a few messages the memory is faulty and the character changes personality. I am willing to pay for a good chatbot but crushon.ai is very expensive. Do you have any suggestions for me?
- Feb 29, 2020
- 144
- 422
if you have the hardware the answer is always going to be running something like kobold locally for a variety of reasons
i've been using koboldcpp being fed into sillytavern running off of a crunchy 2017 i7 and an amd gpu to acceptable results
i've been using koboldcpp being fed into sillytavern running off of a crunchy 2017 i7 and an amd gpu to acceptable results
venus.chub.ai has been pretty good. I don't think the 5 a month is pricy and I've gotten a lot of value for it already, haven't run out yet and used it a decent amount for the last week. I use their mercury plan, I'd avoid OpenAI API since it can get you banned from OpenAI
chai has 70 free messages every 3 hours. HiWaifu has energy system but you can watch 20 ads per hour, use referral code etc to save up energy quite easily. Both have character creation and no nsfw filter. Also janitorai is completely free to use, has character creation and no nsfw filter. all quite fast and satisfying.
Hey, i have r7 5700x and 4070ti.Can i make this kobold thing? And how... I realy dont know where to start
- Feb 29, 2020
- 144
- 422
if you just want to run a language model for coombots,
You must be registered to see the links
goes over most of the relevant information
You must be registered to see the links
but for this purpose we only care about 1:09-5:18 or soyou can stop here, but you can also route it through
You must be registered to see the links
for a better ui and additional features, but do be warned that this can be a much more involved processedit: i should probably go over what those "additional features" bring the table, which is especially important for the context of a chatbot thread
the biggest boon to using sillytavern is its ability to easily import character cards; if you wanted to bang cynthia from pokemon, instead of needing to manually tell the ai everything,
You must be registered to see the links
you can naturally also create your own character cards, but this can be a lengthy process involving a lot of trial and error and i would only recommend this if you can't find a card of your preferred flavor of body pillow or are into some real wack shit
Last edited:
- Nov 30, 2022
- 101
- 147
dude i think we need something like NVIDIA A100 80GB VRAM GPU for 20k $ to run decent 65B model locally
damn i was going to try candy.ai
But i was sceptical as all those "ai" sites that emphasize image generation over realistic chat with chars do not inspire confidence. I mean it looks like they hide weak ai behind shiny bots images. Looks like i was rightCan anybody recommend c.ai/chatgpt4 level chatbots no matter paid or free?
Last edited:
Crushon.ai is also an option that isn't too egregious. Shares some interface with Character.ai if that's your thing.
I don't like the conversation when the chat suddenly loses memory and the character's personality changes, you know? I'm also looking for similar AI sex chatbots. A lot of folks are recommending Soulfun.ai, and I even got a membership, but man, their voice feature really needs some improvement. But gotta say, their chat experience and picture options are really cool.
Last edited:
Soulfun is a complete scam. I fell for It. I hope by leaving this warning I at least save someone.
It has been quite sometime since I replied to this thread. I did find what I was looking for (erotic chatbot rp). Even though the developers do not really like that their work is mainly being used for porn I think It is the best option out there. It is called backyard AI. Easy to use. Cheap if you need cloud GPU (if pc is not powerful enough to run big models locally).
If you are more a handy type, there is also a more crafty method to get nice chats with heavy and light models: oobabooga. It is a public web app that you can download and link with whichever hugging face model you wanna play with. Because my pc aint the best for the big 70B models I found a tutorial to run oobabooga with heavy models on lambdalabs services.
A piece of information before you sail away, context window has no work around. Every model will forget about your chat and turn unreliable after the token (memory) cap is exceeded.
I am available if any of you want a bit more of guide for any of these 2 solutions I have mentioned. I also have some sweet cards for Backyard. Although they are either medieval/roman scenarios OR strong femdom.
Well that is It for me brothers. Cheers.
If you are more a handy type, there is also a more crafty method to get nice chats with heavy and light models: oobabooga. It is a public web app that you can download and link with whichever hugging face model you wanna play with. Because my pc aint the best for the big 70B models I found a tutorial to run oobabooga with heavy models on lambdalabs services.
You must be registered to see the links
A piece of information before you sail away, context window has no work around. Every model will forget about your chat and turn unreliable after the token (memory) cap is exceeded.
I am available if any of you want a bit more of guide for any of these 2 solutions I have mentioned. I also have some sweet cards for Backyard. Although they are either medieval/roman scenarios OR strong femdom.
Well that is It for me brothers. Cheers.
- Sep 19, 2024
- 166
- 99
I'm trying it out at the moment. Not bad for relatively short interactions, which mine tend to be. I do have enough memory for some of the models, but I'll be increasing it soon for the LLMs. What would you suggest? Selecting an LLM and using that to do the deed, or instead selecting a lighter but purpose-built model for NSFW?
I'd say uncensored rp tuned models (>10B) work *alright* but after around 10 to 20 chats you start to predict their outcomes. My gut tells me that even if they are uncensored they were never fed a relatively big amount of nsfw literature. For some reason most of those models are fucking obsessed with nipples. I have even started to ban that word in the character cards haha.
So, with that in mind, I feel somewhat confident to recommend not limiting yourself to a couple of llms. Exploring and changing is much more fun, refreshing and surprising. In my case, I found out I rather smaller models (from 8B to <30) and expanding token context window with that extra memory you are saving.
- Sep 19, 2024
- 166
- 99
Thanks for the reply. I guess a balance between context memory and natural responses is the way to go. I've used other AI bots so I'm familiar with the fact that they're not gonna retain details for very long, and that they'll say some wacky-ass shit every once in a while. Many years of fapping with one hand and slogging through buggy, broken v-0.000.0.01 pre-alpha grindfests has prepared my patience well. So that part thankfully doesn't bother me too much.
I'm using a 7B NSFW model at the moment, by the way. I'd like to triple that and see how my results fare, but it'll be a week or so before I can report on that. Haven't encountered the fixation issues yet, but I've only just started messing around with it. You may be right that the models don't tend to have a lot of NSFW training, but that's what we're here for
- Sep 19, 2024
- 166
- 99
With 64GB of 3200 RAM (really it's running at 2200 or so), I can't even load some 70B models into memory. While the steady use space is around 30-45GB, and I have spare memory available for that task, it seems that when loading some models initially, the memory usage is 50%-100% higher for a moment.
So I've been using an 8x7B model. It's fairly slow, running at about 0.5 seconds per word, and depending on the model, you get non-sequiturs frequently. Hell, the one I'm using now is the thirstiest thing I've ever seen lmao. It'll take some fine tuning for sure, but I'll try to stick to the under 30B models as suggested and see how that goes.
You need ram to do the loading process (from disk to ram or gpu). Wild guess here but maybe you are running out of ram to load onto ram which is actually funny xD There might be an option to solve that tho. Everytime I try to get my hands on inference* I get lost on so many damn options flags and motherfucking modules and versions. As a femdom connoisseur I believe I have the mastery and authority required to accuse all those fuckers of sadistic balls torturers.
*less user friendly ui's with vast.ai rentings.