My ai erotica story generator

[X1000]

This my ai erotica stories generator on miniapps

Free uncensored, try it

Uncensored Muse · Free AI Chatbot

Chat with Uncensored Muse. Unrestricted AI writer for explicit, uncensored long-form erotic novels. Dark, bold, sexual tone.. Try this and more free AI and ChatGPT tools and chatbots on miniapps.ai!

miniapps.ai

If you have advice or tricks to edit it, please tell me

 

[yano2mch]

What model of LLM are you using?

I've personally had some halfway decent results with a few models, but context after a certain point just borks the output.

 

[X1000]

What model of LLM are you using?

I've personally had some halfway decent results with a few models, but context after a certain point just borks the output.

For now for erotica and uncensored topics Grok is the best

 

[yano2mch]

For now for erotica and uncensored topics Grok is the best

While it may be the best, i can't run it myself/locally... Unless there's some model i'm not aware of.

 

[X1000]

While it may be the best, i can't run it myself/locally... Unless there's some model i'm not aware of.

I haven't had local experiences yet ,i think local models more open and easy to train ?

 

[yano2mch]

I haven't had local experiences yet ,i think local models more open and easy to train ?

Well at the very least they are private since your prompt and result aren't saved offsite; Though since you're running on your own hardware, unless you have a GPU, running via CPU can be very slow, and you need at least as much memory as the model size (a 3B is 3 Billion nodes, so 3Gb for the model, possibly more on a GPU so it runs faster as they have multiple layers the pass through to get results)

So far i've been toying with oobabooga's text-generation-webui, plus models from huggingface, a low entry model that's decent seems to be the llama 3B uncensored instruct.

A major problem, maybe that's resolved with the 26B and 70B+ versions (but are way too large for me to consider with only 32Gb ram, we'll see after i have 128Gb); Is context windows and output gets a little off after 5000+ tokens (or about 20k text). A lot of models also only have a context window of 4k or 8k tokens.

As for training, no clue yet, haven't gotten that far. I assume you take a base model that is decent and then give it text and it changes the weights for the next text. like anything NCNN based it's 'throw it at the wall hundreds of thousands of times until you start getting decent results'. Lazy un-debuggable programming.

 

[X1000]

Well at the very least they are private since your prompt and result aren't saved offsite; Though since you're running on your own hardware, unless you have a GPU, running via CPU can be very slow, and you need at least as much memory as the model size (a 3B is 3 Billion nodes, so 3Gb for the model, possibly more on a GPU so it runs faster as they have multiple layers the pass through to get results)

So far i've been toying with oobabooga's text-generation-webui, plus models from huggingface, a low entry model that's decent seems to be the llama 3B uncensored instruct.

A major problem, maybe that's resolved with the 26B and 70B+ versions (but are way too large for me to consider with only 32Gb ram, we'll see after i have 128Gb); Is context windows and output gets a little off after 5000+ tokens (or about 20k text). A lot of models also only have a context window of 4k or 8k tokens.

As for training, no clue yet, haven't gotten that far. I assume you take a base model that is decent and then give it text and it changes the weights for the next text. like anything NCNN based it's 'throw it at the wall hundreds of thousands of times until you start getting decent results'. Lazy un-debuggable programming.

Trying local models for me is a dream now ,i need better pc ,i have only 12 gb ram with built-in gpu

 

[yano2mch]

Trying local models for me is a dream now ,i need better pc ,i have only 12 gb ram with built-in gpu

You can likely still run a local one, but i wouldn't go beyond 1Gb. Mobile versions (intended for phones) i've seen 70-300Mb, yeah they aren't going to be the best for complex roleplaying, but might still be useful. Or you could offload the work to another system, a mini-PC, NAS or other device that has sufficient ram for the smaller model and then just... log in through the interface.

Though with the big push for AI, i won't be surprised if we get very decent chips in the near future letting us just have an AI-USB attachment, PCMCIA (if those are even a thing still...) or card or mini-box (like how printers or NAS you just plug into the network and... poof...) just for doing said processing. I wouldn't expect much til 2028 at least though for them to be designed, tested, and made in sufficient quantities to be available at Amazon or Walmart.

 

[X1000]

You can likely still run a local one, but i wouldn't go beyond 1Gb. Mobile versions (intended for phones) i've seen 70-300Mb, yeah they aren't going to be the best for complex roleplaying, but might still be useful. Or you could offload the work to another system, a mini-PC, NAS or other device that has sufficient ram for the smaller model and then just... log in through the interface.

Though with the big push for AI, i won't be surprised if we get very decent chips in the near future letting us just have an AI-USB attachment, PCMCIA (if those are even a thing still...) or card or mini-box (like how printers or NAS you just plug into the network and... poof...) just for doing said processing. I wouldn't expect much til 2028 at least though for them to be designed, tested, and made in sufficient quantities to be available at Amazon or Walmart.

Thanks for your advice, I'm waiting for new laptop to test and train the local models, i tried the local on mobile found it slow and useless