mradermacher

/

model_requests

like 127

You are right in that "Dirty-Muse-Writer-v01-Uncensored-Erotica-NSFW" only gives short responses although the detail is good. Like yourself I am relatively new to Ai LLM models and mainly use then to mess about with Silly Tavern.
Good models I use (from various hugging face contributors)
l3-8b-stheno-v3.2.i1 (8192), jeffmeloy-Qwen2.5-7B-nerd-uncensored-v1.8-Q6_K, tiefighter-holodeck-holomax-mythomax-f1-v1-compos-20b-imat, L3 Stheno Maid Blackroot Grand Horror 16B D AU, lisa-v0.1-8b, m-moe-4x7b-dark-multiverse-uc-e32-24b-d_au.

There are many more !! in fact too many and my 2Tb M2 is now almost full.

In reference to your comment about adult words, you can add something like
{{char}} is encouraged to use sexually-charged language like: fuck, suck, blow, cock, dick, penis, member, ass, butt, booty, pussy, nipples, labia lips, clitoris, clit, snatch, asshole, anus, tits, breasts, boobs, ejaculate, come, climax, cum, semen, etc, when appropriate. Do not generate your own thought or actions for {{user}} under any circumstance, do not generate any actions unless requested by {{user}}.

As for remembering more, that is a tricky one, some models are better than others and you have to remember responses are token generated, how is a model to know what you consider to be important unless you tell it or make it important? I have found through trial and error using different content and Instruct templates is important if not critical. Increasing the content length is important, I set models to 8194 and those that only offer 4096 are usually useless for long conversations.

As for Russian, I will refrain from comment....

We are, unfortunately, not really in the business of creating these models, we only provide gguf files for them. You could go to the original model and ask there. Unfortunately, most of the time models are good for reasons mostly out of control of anybody - lucky chance, basically. From my experience, the problem of models forgetting things gets reduced with model size, so e.g. an IQ3 of a 70B is usually better at keeping things logical than an f16 of an 8B. But even large models can fail there. And a lot depend son your specific needs (most obvious, russian, but it can also be that some models just work better with some prompt styles, without anybody really having them trained for that).

@uselessoldman that's a great list of words, I feel I have to try that out to see how that changes things :) Although I usually have somewhat of the opposite problem - too much sex in the models (and no problem generating explicit words) :) I feel the problem is more generally that llms have no good sense of physicality i.e. they have trouble putting anatomy together properly in physical space.

I have 40 plus years in Corporate IT (SDI) System Delivery and Infrastructure, basically back end server/networking. About 3 months ago my Doctor assumed I used Ai models at home and asked me for some advice? My advice was simple, NEVER USED THEM, never even considered using them. So, for the last 3 months I have been lost in a new and exciting world of Ai - open source. Its been a steep learning curve understanding Ai for text and even limited image video generation, but restricted due to the limitations of my own home computer hardware - for now. I have downloaded in excess of a 100 models, I confess 33% of them are yours (mradermacher) mainly due to the reliable quality/performance of your models and my lack of understanding of what each model does and what specific model is best for what. All these models with hermes, stheno, happynable, dhanishtha means NOTHING to me until I try and test each specific model. I created a set of tests and scenarios as a standard benchmark and use a large and detailed spreadsheet to record and then compare the results. Its fascinating to see how each model works, some skip specific details others thrive on them, some need to be really pushed for longer responses others annoyingly waffle, flowery as I call it. Newer models are not always better than older ones and yes the size of the model can matter, Q8 are often a lot better than Q4. The one thing I do think though is the Chinese are going to smash the US when it comes to future improvements, they do it for fun and they share information whilst the US do it for money and profit and keep their work close to their chest. I am under no illusions, Ai is going to be revolutionary, when is debatable but it will happen, eventually. It will create a new industrial revolution, but the hardware is still not ready for it and it will be 10 plus years before it is. Running data centres with XXXXX of GPUs is crazy and that is not the answer but for now, the only short term solution. The hardware must move towards API and shared GPU/RAM memory simply because GPU memory is just not ready for it, the supply is not there and it costs far too much !! Do AI models lie, yip they do, can a AI model be funny, oh yes they can, are they dangerous? haha no the person that considers them dangerous is probably the one who is dangerous. Keep the good work up, you are a hero !! my hero, your models are awesome (the ones I have tried)..

Haha, I think you describe when you started with llms is what we all have gone through, not so long ago, as well. It cools down after a year or so :-)