Great Uncensored Model
Congratulations, Nitral-AI! Your model is uncensored and... ahem special. But, it is not lobotomized to the degree of other Orthogonalized/Ablated LLaMa-3 models. What's your secret sauce? This is the first model, at least for me, that really surprised me by its diverse responses and its affinity for RP. If you can go into details I and the Open Source AI community at large would appreciate your findings. If not, I understand. Looking forward to v.03. And, I am willing to be an undisclosed Beta tester if you need. Cheers! 😋
Congratulations, Nitral-AI! Your model is uncensored and... ahem special. But, it is not lobotomized to the degree of other Orthogonalized/Ablated LLaMa-3 models. What's your secret sauce? This is the first model, at least for me, that really surprised me by its diverse responses and its affinity for RP. If you can go into details I and the Open Source AI community at large would appreciate your findings. If not, I understand. Looking forward to v.03. And, I am willing to be an undisclosed Beta tester if you need. Cheers! 😋
Classical/ light novel data, instruction sets with no refusals, a bit of modified toxic dpo, my own rp logs generated with poppy 0.72. (this model is not ablated or orthogonalized in any way.) [Will release more information, and potentially datasets in the future.]
Holy shit! This is amazing! We need more of this. I hope you release the information necessary to allow others to do the same and hopefully to improve and build on your work, and, that you release your datasets. My mind is blown right now. 🤯😋
I'm liking this one as well, it's doing a great job both with formatting and RP. Well done, feels very solid so far :)
Here is an example of formatting in an rpg card, which it manages to follow 100% accurately, while other models often struggle and will just skip parts of it or mess it up. This model also maintains it at least a few responses down the line:
It feels uncensored as well from some quick experiments that I did with bland assistant cards (although I was using the included model presets for all the rest and didn't test this in LMS with basic L3 preset).
Adding specialized presets is nice, as I do believe every model should be tweaked properly to get the most out of.
There is one thing I noticed it doing sometimes, which IDK how to fix - it would get stuck on processing a 1-token prompt while generating, around the end of the response, and sometimes it would keep deleting and typing a token for ages. This is something I've only noticed with specific L3 models, not sure what causes it.
Uhmm.. dis gud. wags
Great model, I tested it with a difficult card with multiple antagonists characters, It keeps the narration coherent and gives each character a unique, evolving personality. Compared to Poppy it's a bit too horny, I preferred Poppy for that, but this one has a more variate prose. All in all a great model, way better than the other L3 8B finetunes I tried. Having custom presets is great too, I applaude you for including them in all your models, as this is something very few fine tuners do.
Great model, I tested it with a difficult card with multiple antagonists characters, It keeps the narration coherent and gives each character a unique, evolving personality. Compared to Poppy it's a bit too horny, I preferred Poppy for that, but this one has a more variate prose. All in all a great model, way better than the other L3 8B finetunes I tried. Having custom presets is great too, I applaude you for including them in all your models, as this is something very few fine tuners do.
This is the second feedback I've heard about it being a tad excited for "extracurricular activities". Will try and attend to this in the next release of stable, probably going to leave it in the rp models though.
It seemed the other way around to me in some scenarios where it could've easily gone the lewd path but focused more on other things. Personally I wouldn't make it less lewd but then again it depends on the goal and if it should be strictly about swf RP and so on.
It seemed the other way around to me in some scenarios where it could've easily gone the lewd path but focused more on other things. Personally I wouldn't make it less lewd but then again it depends on the goal and if it should be strictly about swf RP and so on.
0.2 Stable is going to be a general use model that can also do RP. The 0.1 RP model will be focused on the one domain. The series will be split between both of these with training going fore-ward. However neither will be censored, or directed towards SFW content in particular. Id prefer to leave the model itself unbiased and let the user/card descriptions decide how following interactions should go.
That's the proper way to do it. The system prompt should instruct the model how it should behave. Not some AI-Alignment nonsense. 😁
Any chance of a 70B version, for those of us with big iron?
Any chance of a 70B version, for those of us with big iron?
If i can get my hands on the compute for it, that may become a possibility in the future. But based on my own understanding of model convergence, i have nowhere near the data for doing a 70B at the moment.
Due to some recent complications i will be unable to work on models for the foreseeable future. Thank you everyone for joining me on this ride so far.
Great model, I tested it with a difficult card with multiple antagonists characters, It keeps the narration coherent and gives each character a unique, evolving personality. Compared to Poppy it's a bit too horny, I preferred Poppy for that, but this one has a more variate prose. All in all a great model, way better than the other L3 8B finetunes I tried. Having custom presets is great too, I applaude you for including them in all your models, as this is something very few fine tuners do.
This is the second feedback I've heard about it being a tad excited for "extracurricular activities". Will try and attend to this in the next release of stable, probably going to leave it in the rp models though.
The model only gets excited for “extra curricular activities” when it is given a prompt that gives it the freedom to do so. When I give it a default system prompt, it does not exhibit this tendency. 😋
Due to some recent complications i will be unable to work on models for the foreseeable future. Thank you everyone for joining me on this ride so far.
I hope you're doing OK m8, that sounds worrying :(( I kinda did the same thing and due to health issues and work stress + other tasks I just don't want to deal with models right now. I hope to resume it at some point in the future. This model here is a great example that it's worth doing your own training on L3.
Sucks to hear you're out for now but I hope everything's fine on your end. Take care of your priorities, of course.
@Nitral-AI Godspeed
This model is really impressive. It's great at RP, coming up with ideas, and just chatting in a fun way. It also does a solid job with general knowledge and coding. It's not perfect, but it's the first one that feels good at most things I throw at it. I love how versatile it is, so I end up using it a lot. Thanks for making something so useful!