This seems better than Hexo
Nice work here. I grabbed Hexo first and then this one just for fun, but finding much more verbose, detailed, original and generally smarter results from Nab than from the newer one. Suggest you don't abandon this!
I take it back, Hexo is pretty incredible in the right context. Nice work on both.
Yeah Hexo is interesting to me in that its design was inspired by taking the general approach used in Naberius and scaling it up. Shame is I don't have Hexos original design notes as in, the specific selected models, pefts, pairings, and merges of pairings - it's six funetuned models merged, but one twist is for every two merged a lora that thematically compliments that pairing is hand selected, fused, and the idea was applying low rank adapters after a pair merge may serve as a common ground to smooth out and extreme differences between them. I had an entire page in 8pt font, 4k screen in Notepad++ absolutely filled with hours worth of candidate models, presumed optimal pairings, and basically all this was assembled in a six way merge pattern done piece by hand picked piece.
And I fucking sold the PC which had all those notes and theories behind the work before it got summarized on the model card 😆. Hard times had to eat. Glad this space changes so fast, losing a lot of work feels bad one day, then the landscape is so different it's already time to move on and try something completely different.
My personal takeaway is Naberius is stable and efficient, Hexo is complex and occasionally interestingly lucid, then inane the next moment. Touchiest model I've ever endlessly tweaked sampler settings on ever. There's a cool person on here somewhere that fused the LIMARP Lora to Hexo and has it up in ggml format, personally I think he was on to something. Ironically one of Hexos low level merges already has LIMARP in it lol. Pour it all over a stacked mess like syrup and nobody notices the merge was a failure 👀 [I unironically use their ggml of hexo all the time, just works better]