Request Tool KoboldAI + Erebus Model for text-based adventure

Lakius

Member
Mar 22, 2019
158
650
ok lets say I put it in "Memory" what should i write?
not wearing panties?
I've added this but still it seems they grow back
Use World Info. Key(s) is where you put their name, What To Remember is where you put the description. Bear in mind this is best used for what they look like at scene start. Update this after your scene is done, whenever something important changes that would affect a new scene. (haircut, recently proposed instead of dating, no longer a virgin, etc) Avoid details that are irrelevant, like an estranged sibling you never want to actually bring up in the story.


Use Memory for a summary, it goes at the very top so it doesn't affect the output much. This is like the summary blurb on the back of the book. Good for overall tone, weaker for scene specifics. I highly recommend this for any story longer than one scene.
Use Author's Note in Erebus for genre tags. These are pretty strong and I change these often, even mid-scene. This is inserted right before the last action, so it can be a somewhat useful bludgeon to get the AI to co-operate if you instead use it for scene reminders. It's way easier to just write a guiding sentence that would be shit storytelling for a real book, then delete it once the AI acts on it.
Use World Info for characters or other important things. These go after Memory and before the action history, so your scene can still take precedence. If you are new, just pin everything you want, and try to keep entries short and sweet (150 tokens or less with Max Tokens at 2048), and/or increase your Max Tokens in settings. For new characters, get your AI to describe them mid-scene, then just copy and paste that into the World Info record with their name, maybe adding some character traits like pensive, dramatic, seductive, etc. I try to avoid clothes because then I need to manage the entry if it changes often, but it's good for characters with an outfit they're known for, like maids or soldiers or that one character who always wears shorts.

You need to remember that the AI has a limited memory. If you write that a character wears a bikini to a boardroom meeting, you'd better put it in World Info or womewhat regularly bring it up, or else they'll be wearing something boardroom appropriate later. How soon the AI will forget is based on Max Tokens minus your loaded Memory + Author's Note + World Info. Check your server log for a better idea, it will show you exactly how much history the AI is working with.

For clothing, yeah the AI is as dumb as a brick, but it's often the memory issue. Just delete any lines that don't make sense. You really do need to delete a lot, don't just hit submit again and hope it gets better. Write out your own lines if the AI just isn't going where you want, even half a sentence is good enough most times. We're just not at the level where an AI can write a good book on its own, and this is mean to point out, but if you give it bad inputs it will give you bad outputs.

For positioning, create a world info entry for the room they are in and pin it. Again, keep it short and basic, just saying room type and key furniture. You should try to remove any ambiguity when characters move about. If a character isn't given a position and the scene just started, the AI will probably want to put them somewhere. In the middle of a scene, this tends to not be as important.

Repetition corrupts, leading to more repetition until it gets to a looping state. If you start seeing a pattern in your outputs, break it ASAP, and take a look for where it started to fix that too. This can be as obvious as one line repeating over and over, or as subtle as a conversation turning into "[Name] frowns/smiles/laughs/etc. '[Simple reply]'" repeated over and over with the AI rendered incapable of escaping or writing any variety by the 4th or 5th time it appears. It's my biggest gripe, because most real authors try to avoid structural repetition like that. Your character moving about constantly is likely this issue. By the third time they did it, the AI was thinking "Alright, this is the pattern, I predict it's going to happen again." Even though it reads terribly and doesn't make sense.
 

DevonDekhran

Member
Aug 27, 2017
104
111
Since I'm far too new to this all, and was just using colab for the kobold models, is it even possible to run "well" the newer erebus-6.7B between a 5900x, 3070 and 32gb of RAM?
The CPU is getting here somewhere around this week and I wanted to know where to put my hopes at. Seeing the GPUs you guys are using already set them quite low though
 
Last edited:

Dir.Fred

Member
Sep 20, 2021
141
246
It's the 8GB on that 3070 that's your limiting factor. Nothing matters anywhere near as much as how many layers you can unleash your GPU's CUDA cores on and that's all about VRAM. Any CPU embarrasses itself when pretending to be a GPU. :geek:
 

DevonDekhran

Member
Aug 27, 2017
104
111
It's the 8GB on that 3070 that's your limiting factor. Nothing matters anywhere near as much as how many layers you can unleash your GPU's CUDA cores on and that's all about VRAM. Any CPU embarrasses itself when pretending to be a GPU. :geek:
I understand, just wanted to guess how far would the suffering go. But somehow this thing got here earlier than expected so I'll be testing it tonight. I hope it ain't too far above 30s per gen

Edit: Welp, 45seconds. Pretty good considering my last cpu attempt. But I'm definitely hopping for gpu prices to get close to affordable at some point in time.
 
Last edited:

carter111

Newbie
Jul 26, 2017
48
154
can you share some of the stories? I wanna see if it's worth the hassle to set it up.. especially cuz I already have stable diffusion.. ANY FETISH is ok (except guro, snuff) I wont judge I am only interested in seeing how coherent the AI is, and if it can remember the genders and names of the characters... (I set up a free browser AI with 2 girls and a dude and even tho I set their genders right it kept going back and forth addressing one oh the females as he or she randomly)
 

Lakius

Member
Mar 22, 2019
158
650
can you share some of the stories? I wanna see if it's worth the hassle to set it up.. especially cuz I already have stable diffusion.. ANY FETISH is ok (except guro, snuff) I wont judge I am only interested in seeing how coherent the AI is, and if it can remember the genders and names of the characters... (I set up a free browser AI with 2 girls and a dude and even tho I set their genders right it kept going back and forth addressing one oh the females as he or she randomly)
It's much simpler than SD to set up locally. I'm not sure what web version you've tried (much more variety than image ones), but coherency for natural language models usually all about context. Taking the time to create world info entries and quickly turning new characters into them is key. Also, model choice and size helps.

If you're keen on reading something, feel free to PM me directly. A lot of my stories have a lot of editing and steering done, but I can send you some portions that were written on a tablet. There's a lot more context and nuance I'll leave unsaid, unless you want to hear more in a PM. I've put enough walls of text here in this thread. :p
 
  • Like
Reactions: smirk

Lakius

Member
Mar 22, 2019
158
650
I understand, just wanted to guess how far would the suffering go. But somehow this thing got here earlier than expected so I'll be testing it tonight. I hope it ain't too far above 30s per gen

Edit: Welp, 45seconds. Pretty good considering my last cpu attempt. But I'm definitely hopping for gpu prices to get close to affordable at some point in time.
Just to confirm, you are leaving the Disk cache slider empty, right? Even the fastest SSD on the market is glacial compared to RAM, and the sliders on KoboldAI are incredibly misleading for newbies.
 

DevonDekhran

Member
Aug 27, 2017
104
111
Just to confirm, you are leaving the Disk cache slider empty, right? Even the fastest SSD on the market is glacial compared to RAM, and the sliders on KoboldAI are incredibly misleading for newbies.
Indeed. I just used GPU+CPU, had tried the cache before I knew what it was and I just gave up after 2 minutes had passed without anything happening at all. But it was on an nvme so yeah, quite horrifying
 

Lakius

Member
Mar 22, 2019
158
650
Indeed. I just used GPU+CPU, had tried the cache before I knew what it was and I just gave up after 2 minutes had passed without anything happening at all. But it was on an nvme so yeah, quite horrifying
Well, you can always load up a smaller model and give it a shot. Even the larger ones are an assistant at best.
 

mhg87

Newbie
Jul 14, 2018
60
209
I think that the comments regarding the quality of output from AI are correct. It's still mostly of below average for free lewd/porn writers due to its lack of ability to focus or have common sense or understand timing. It'll rush through a sex scene in a couple of sentences with barely any detail and then have you doing errands at the store for paragraphs. Or it'll forget stuff and start going in loops and being surreal. You've got to wrestle with it to keep it on track constantly. The only reason it is worth doing is either if you are a writer and super stuck for ideas and want to take a shot in the dark with this as you've no one else to bounce ideas off of or you're looking for some fap material like most of us. Not many people are willing to supply a stream of on-demand, possibly niche, fap material when you're feeling horny and do so rapidly. For that kind of thing it's pretty good as if you're feeling horny you'll overlook all but the most absurd, like it writing characters with too many body parts or bizarre mutations or accidental comedy and it might be difficult to correct and get back on track.

There's no doubt that right now the average human writer would make you something vastly better, especially if they had the same or understood your preferences well. Kinks and details are so important in writing lewds but what details you focus on really matter. Do you want detailed information on sweat dripping down abs or the way a pair or breasts heave and jiggle? A good writer and/or one that knows your interests well of course will blow it out of the water, no doubt.

The real story experience in writing actually happens in your head, not the page or screen, which is why it needs to feed the head with the imagery that matters. I'm sure that at some point you'll have easy tags to tell it to expand certain things and have it learn more variety and focus on your interests but so far it's not there yet.

The two things it seems best for is dungeon crawling and adventures or lews. Honestly the latter especially as there's obvious barriers to that with humans, while arranging games of RPGs isn't too difficult and you can schedule that while horniness can be more of an "in the moment" thing. Yet the crazy thing is that they're not all diving more into supporting lewd content and promising privacy, a combination that would drive customer base for sure. The porn/lewd industry is pretty big, after all.
Hah, that's where I'm at right now too. Every time I use Kobold there's a little voice of reason in the back of my head saying, "Why the fuck aren't you just writing porn yourself, you idiot?" I mean, with all the adjustments I have to make to the AI's output, I basically am.
 
  • Like
Reactions: smirk

DevonDekhran

Member
Aug 27, 2017
104
111
Hah, that's where I'm at right now too. Every time I use Kobold there's a little voice of reason in the back of my head saying, "Why the fuck aren't you just writing porn yourself, you idiot?" I mean, with all the adjustments I have to make to the AI's output, I basically am.
That's why I prefer the chatbots for this purpose, but in general they also aren't too much better than this. Specially considering the heavy censorship on proprietary ones like replika, and the lack of options for self hosting. At least I only know of pygmalion for now
 

smirk

Member
Jul 2, 2019
145
197
Been playing around with lately, and had quite good results. It's a weird chatbot/rpg adventure hybrid thing using KoboldAI as a backend. Using the Pygmalion 2.7B model and some of the pre-made characters, it can be a fun for a while and actually produces decent results. Can recommend if you're into that sort of stuff.
 

cooperdk

Engaged Member
Jul 23, 2017
3,375
4,902
Anyone figured out how to make novelai modules usable as softprompts in kobold or oogabooga?
They are trained differently so I am sure that is impossible. Kobold softprompts are binary numpy objects and a json file. I am not even sure what format NovelAI uses but in the softprompt data file, it is a JSON with binary data, descriptive text (like the meta.json in Kobold softprompts) and a list of loss data.
I can see how it would make sense to convert them.

But another point is that the NovelAI modules are trained for the NovelAI model which is not available (it is crazy big, like much larger than fx Pygmalion-6B, and we're probably talking 32 or 64 GiB). I do have the dataset used to train it and it requires a rack of TPUs to even get somewhere. It is basically thousands of books. 2 GiB worth of txt files to be precise. Post apocalyptic novels alone are 38 full books and artificial intelligence are 48 books, some of which are actually 5 volumes in one text file.
So it would be necessary to retrain the softprompts anyway.

Unfortunately, it is very difficult at this time because Colab TPUs don't work and the online trainer requires TPUs. And the offline version available with koboldai (tuner.py in the root folder) is not working, at least not in Windows. I actually have some collected datasets for doing softprompts but ... yeah well.
It would be hella nice if KoboldAI had a tuning tab just like the SD Web UI (where it's called hypernetworks and embeddings but with each their purpose).
 

cooperdk

Engaged Member
Jul 23, 2017
3,375
4,902
BTW, Pygmalion-2.7B or Pygmalion-6B are a lot better than erebus, especially for NSFW. Pygmalion is supposedly close to NovelAI and is supposedly trained on some of the same data. It is big, the 2.7B is about 6 GiB and the 6B is 16 Gib but split in two so it can run on as "little" as 10 GB which is the requirement for the largest file in the set.


 
  • Like
Reactions: Dir.Fred

smirk

Member
Jul 2, 2019
145
197
BTW, Pygmalion-2.7B or Pygmalion-6B are a lot better than erebus, especially for NSFW.
Pygmalion is also tuned specifically for chat; have you been able to make it work for stories/adventures? I've given it a few shots but it inevitably goes into chatbot-mode after a few prompts. It is exceptional for use with TavernAI and the likes though.
 

poppel

Newbie
Jan 28, 2019
15
7
I can't start it. Everytime I try to start it, using the play.bat, it says that the system can't find the stated file.
 

Venca

New Member
Aug 10, 2017
3
18
Have you guys tried generating adventures with Meta's leaked llama.cpp. I've started last week and am having a blast.

Basically, you download koboldcpp from here and follow what's written in the instructions file (you'll need to launch a few scripts). Note this is NOT KoboldAi in itself, it merely reuses KoboldAi's GUI to make working with llama.cpp more user friendly, it will not load KoboldAi's models.

Then you grab a model from here . Llama is the model that was originally leaked; Alpaca is the Llama model, but more trained by users; Vicuna is the Llama model, but trained on ChatGPT's responses, it is meant to be similar to ChatGPT, and is more censored because of this.
Basically, grab Alpaca for now, it's what I'm using. You want models for llama.cpp, so ggml or gptq. Here's a link for the download of a model you can use (13B version, so you'll need ~13 GB of RAM, smaller models can be found in the link on reddit).
They used to have torrent link for these models, but recently removed them for some reason.

You then launch koboldcpp.exe, it will prompt you to select a model, select the .bin file you've downloaded. It will load the model (can take a minute, also might require permission from your firewall) and give you an address to open in your browser (it's a local connection, so everything happens on your PC), and you're basically ready to go. Choose "Scenarios", and "New Adventure". The adventures are not NSFW by default, you have to make the AI understand you want it to go in that direction by wording your prompts accordingly (or add NSFW descriptions in World Info). As I said, I've been using it this last week and am having a blast, though I didn't have much experience with AI adventures before, so YMMV.

For completeness sake, posting the link to the original guide I followed (in Russian)
 
  • Like
Reactions: defp and Dir.Fred