Just wanted to say THANK YOU
Hi, Mindrage. I registered just so I can say thank you for your work on the amazing model that you have posted. I have been engaged in the AI for over a year now, and I think I've seen it all and tried them all at . Your model has remained my absolute favorite - it is fast, it is uncensored (can eloquently present both sides of any argument, which is so rare these days), is very intelligent (like you said, seems very much alive often), it even has a funny personality side to it (which is very unusual and brought a lot of fun). I don't know how you pulled this magic off, I just want to say THANK YOU for bringing this model to life. Your profile is permanently open in my browser tab, where I refresh it often in hopes of finding more of your work. Again, big thanks!!
Thank you so much man, I'm truly happy that others enjoy this model, too (still!).
But again, I really have to point out that the work I did was the least part here. I listed and linked where credit really is due in the model card at the very beginning.
My contribution was mostly just realizing that this model had unbelievably been overlooked and that there didn't exist a quantization at the time, so I made one to help make it more available on user hardware.
Strangely, I still haven't found a model that (for my use case, which is "general assistant/chatbot") could replace it, not even the new Llama2 models.
Mythomax (llama2) seemed close, but comes off as pretty "soulless", like all the llama2 variants I have tested so far.
If I DO find what I believe to be a superior one, I will upload or at least link it, of course.
If not, I might have to start working on recreating the original as faithfully as possible with a Llama2 13B model. It should be possible, since most of what made it special is available as LORAs I think, so with a bit of luck some trial and error could lead to an approximation of the original "recipe"...
Yep, I realize most of the credit goes to the authors of the source models, but your quantization somehow made a big difference. I tried all the parents, and Llama2 variants as well, and none come close to the 'personality' and 'creativity' of this model. So I am also sticking to it for now. Mythomax is good but not as good. Vicuna 1.5 (llama 2) is a little better than Mythomax in my opinion, but also not as good.
I have a certain way of testing censorship, and there is a well-known person here who quantisizes many, many models that come out, and does it quickly too. I noticed a strange thing, that this person's models come out more censored and 'aligned' than same models quantisized by other people. It could be that this person is using wikipedia dataset as part of the quantization (not sure, but wiki is very censored in my opinion and is not good for this purpose), or something else is going on. So whatever you did, your variant works perfectly and is very unbiased (which makes it more eloquent and expressive answering all sorts of questions and assuming all kind of points of view). If you have time and resource, and could quantisize more models that you like, that would be very much appreciated! Having that one person dominate here as the only/main source of models is maybe not such a good situation to have, especially because, as I mentioned, their quanization seems to make said models have some more censorship or an agenda.
Hey, Mindrage! I hope all is well. Just wanted to say that I compared the outputs of this model with llama2-70b and 30b variants, and it still outperforms even those larger l2 models. This 13B model is like a young Motzart among his peers. Edit: well, the Motzart comparison might not have been appropriate, but it is fascinating how impressive this model is given its smaller size.