They just released the Clio model which is nearly as strong as Llama 7b even though the model is 3b. Currently, it is only available to Opus but in two weeks, it will be released to everyone.
- Sep 3, 2016
- 1,936
- 3,809
They released a graph saying that it was not as strong as Llama 7b but was nearly as strong. Even though it was 3b. According to the chart, it was also stronger than a gpt-neox 20B model and nearly as strong as a gpt 175B model even though it was a 3b model. They used 6 terabytes of training data text for clio.
Last edited:
- Sep 3, 2016
- 1,936
- 3,809
It depends in usage. I don't think Llama 7B or even 30B is as good as Clio in storytelling.
- Aug 7, 2020
- 2
- 3
I have joined the fray and sunk a lot of time into AI chat and "adventure mode" over the past 2 weeks. Pretty cool!
KoboldAI was nice due to the simple install, but after a while I switched to SillyTavern + Agnaistic with the Oobaboga UI to run the AI on, because it lets me load the "WizardLM 13B Uncensored" model on an RTX 3090 (using 8-bit quantization instead of 16-bit floats - something that KoboldAI doesn't support).
It's a lot of fun, the AI really comes up with some very spicy and detailed descriptions for the action, even if unusual fetishes are involved. Only the 2048 token context length (the amount of earlier text that the AI reads to figure out what to write) is a real issue. Like: 1) accept quest in guild, 2) engage in a little romp with my adventurer party companion, 3) nobody even knows about the guild or the quest anymore.
For that reason alone, I subscribed to NovelAI (up to 8192 token context length). The writing is decent, but not as good as WizardLM 13B. Yes, I tried all the presets and many, many re-rolls. In terms of spicy, colorful descriptions of adult and fetish stuff, I have to say, the Open Source product takes the crown.
I've also trained a NovelAI module on a few megabytes of carefully curated beast stories (all involving human females w/dogs, horses and big cats). The module is for NovelAI's "Euterpe" model because "Clio" doesn't support modules yet, but it did improve scenes involving this fetish quite a bit. If you want it, it's attached.
KoboldAI was nice due to the simple install, but after a while I switched to SillyTavern + Agnaistic with the Oobaboga UI to run the AI on, because it lets me load the "WizardLM 13B Uncensored" model on an RTX 3090 (using 8-bit quantization instead of 16-bit floats - something that KoboldAI doesn't support).
It's a lot of fun, the AI really comes up with some very spicy and detailed descriptions for the action, even if unusual fetishes are involved. Only the 2048 token context length (the amount of earlier text that the AI reads to figure out what to write) is a real issue. Like: 1) accept quest in guild, 2) engage in a little romp with my adventurer party companion, 3) nobody even knows about the guild or the quest anymore.
For that reason alone, I subscribed to NovelAI (up to 8192 token context length). The writing is decent, but not as good as WizardLM 13B. Yes, I tried all the presets and many, many re-rolls. In terms of spicy, colorful descriptions of adult and fetish stuff, I have to say, the Open Source product takes the crown.
I've also trained a NovelAI module on a few megabytes of carefully curated beast stories (all involving human females w/dogs, horses and big cats). The module is for NovelAI's "Euterpe" model because "Clio" doesn't support modules yet, but it did improve scenes involving this fetish quite a bit. If you want it, it's attached.
- Mar 3, 2021
- 517
- 946
If by "unlimited input" you mean "let me dump this 300 page book in memory and generate from there", this isn't really possible from a technical level. All AI models work within the confines of a set amount of "tokens", which are a set assortment of words, punctuation marks, letters, or a combination of all three, which serves as its memory. When it hits the limit of it, it will inevitably have to throw some of it out to make way for new information. For most models you'll find, this limit will range from anywhere from 2048 to 8192 tokens, with a few (OpenAI's flagship models and a select assortment of experimental indie ones) reaching into 16K tokens. In practical terms, NovelAI gives an estimate of about 2.9 characters per token, though this may vary depending on the actual contents of the text.
You can get around this with NovelAI, however, by using the story's memory and lorebook entries. The memory will allow you to put the most important information to always be in memory at all times, while the lorebook will allow for relevant information to be brought back into memory whenever specific keywords are hit by the text. You can then focus your generation efforts on a specific scene at a time, and then carry over relevant story information or cool stuff generated by the AI into memory/the relevant lorebook entry.
- Dec 6, 2016
- 82
- 72
Thanks for the explanation, but NovelAI is different than ToolBaz.
When you put the input in NovelAI, it will continue with the storytelling.
In ToolBaz, you set the input of what you want to happen in the story and ToolBaz will create the story from start to finish based on the setting of your input.
A similar one with ToolBaz, which I know, is Ryts.
Unfortunately, Ryts have even far less input than ToolBaz and no spicy slang words either.
So I was wondering if anyone knows any ToolBaz better alternative.
- Aug 7, 2020
- 2
- 3
The entire concept of all the current AI stuff is that a trained neural network is handed the last X words and asked to decide the 1 single next word, repeated eg. 200 times to produce 200 words of output (except instead of words it's tokens - word fragments).
Behind the scenes, that'll be the same for ToolBaz, NovelAI, TalkDirty.ai and everything else out there.
If you have the RAM, you could try self-hosting the mpt-7b-storywriter model with 65536+ input tokens. Its writing didn't impress me, however. As a compromise, there are the WizardLM-13B-SuperHOT-8k models with 8192 input tokens and with much nicer output quality. I tried asking it to write a story and the result was okay, see attached screenshot (it took about 22 seconds to generate on a 3090)
Or simply wait it out. GPT-4 supports 8192 input tokens (and supposedly a variant can do 32768 input tokens), maybe ToolBaz or a competitor with the same service will switch to GPT-4 in the not too far future.
Behind the scenes, that'll be the same for ToolBaz, NovelAI, TalkDirty.ai and everything else out there.
If you have the RAM, you could try self-hosting the mpt-7b-storywriter model with 65536+ input tokens. Its writing didn't impress me, however. As a compromise, there are the WizardLM-13B-SuperHOT-8k models with 8192 input tokens and with much nicer output quality. I tried asking it to write a story and the result was okay, see attached screenshot (it took about 22 seconds to generate on a 3090)
Or simply wait it out. GPT-4 supports 8192 input tokens (and supposedly a variant can do 32768 input tokens), maybe ToolBaz or a competitor with the same service will switch to GPT-4 in the not too far future.
- Dec 6, 2016
- 82
- 72
- Sep 3, 2016
- 1,936
- 3,809
- Dec 6, 2016
- 82
- 72
- Sep 3, 2016
- 1,936
- 3,809
The current model Clio is already the BEST storytelling model ever. But it is only 3B size. While the model mentioned on the screenshot is bigger, fixed and more powerful. And will be released soon.
Any free/one time fee alternatives for AI-powered lewd games (specifically text generated by AI)? I'm all for switching to something different than AI Dungeon (I'm personally not that concerned by the things pointed out in here, but I completely understand why people are) but I don't appreciate NovelAI's lack of free plan whatsoever and only a free trial. Replika.ai feels cringe ("your AI friend who cares", no, no it doesn't) plus the good options are paywalled anyway, character.ai needs jailbreaking, which I don't want to do, I've tried kobold on google colab, it wasn't really that good, don't have the specs to run anything from them locally if there's a better model than what I have tried... I have looked in many places, but all I have come across are some more options, which come with drawbacks of their own, and therefore not necessarily better:
starbyte.ai/ai-anyone/#/ someone's passion project on Reddit, currently the "Behavior" button does not save so all you can do is chat with the default character, but if it gets fixed, then it has the potential to be amazing, and honestly it is already good for a bit until you get bored of Mahuri.
subspacestudios.io/domgpt-brat-princess-ashley/ femdom only, and only that model available for free, rest is paywalled, but seems like something which could be worth paying for if you're into it, since the complete game is actually much more than just text
Virtual Succubus, but that's not really as much of a text game as you'd want, and again, femdom only
I try to look for new AI sites which could be of interest for lewd use (e.g. happyaccidents.ai does NSFW images for free, amongst other things), but the text generator industry seems to be advancing somewhat slowly, which is a shame, because I feel like if you get a good AI text generator for lewd stuff, then you don't really need any other game at all, or at least most of the time. I'll consider NovelAI but I feel iffy about subscriptions so hopefully there will be more comparable alternatives soon for a better price rather than just AI Dungeon
starbyte.ai/ai-anyone/#/ someone's passion project on Reddit, currently the "Behavior" button does not save so all you can do is chat with the default character, but if it gets fixed, then it has the potential to be amazing, and honestly it is already good for a bit until you get bored of Mahuri.
subspacestudios.io/domgpt-brat-princess-ashley/ femdom only, and only that model available for free, rest is paywalled, but seems like something which could be worth paying for if you're into it, since the complete game is actually much more than just text
Virtual Succubus, but that's not really as much of a text game as you'd want, and again, femdom only
I try to look for new AI sites which could be of interest for lewd use (e.g. happyaccidents.ai does NSFW images for free, amongst other things), but the text generator industry seems to be advancing somewhat slowly, which is a shame, because I feel like if you get a good AI text generator for lewd stuff, then you don't really need any other game at all, or at least most of the time. I'll consider NovelAI but I feel iffy about subscriptions so hopefully there will be more comparable alternatives soon for a better price rather than just AI Dungeon
- Nov 15, 2018
- 365
- 400
- Aug 20, 2018
- 3
- 4
New model, Kayra, came out for Opus ($25 dollar tier), and it's coming to other tiers in 2 weeks. Kayra is a 13B parameter model that outperforms GPT-3 Davinci (a 175B parameter model) in every applicable metric. I've tried it out, and it's pretty good. Up next is a new image generation model, then who knows.
- May 8, 2017
- 2,889
- 3,291
Sadly, small models may outperform GPT 3,5 (GPT 3 is not that impressive) on various benchmarks, but when it comes to storytelling, richness of language and general undestanding of commands, they lack a lot. Novel AI masks it with very short outputs but it's still quite clear if you fooled around with non-censored (kinda) version of ChatGPT that it's much weaker model. But it's a great start. Probably the best models I've seen for erotic storytelling purpose as of yet. Would love to see local versions of them. But most of all, I hope for at least GPT 4 quality model with fast outputs. With some huge number of token limit (memory). Claud already has 100k and it is very likely that limit will be functionally lifted soon, so memory won't be an issue anymore. It might take a year or two with how things go.
- Dec 6, 2016
- 82
- 72
- Sep 3, 2016
- 1,936
- 3,809
It's really good. Leagues better than the previous models. It actually remembers thing without having to store it in the lorebook first (though that helps).