Merging Info
You'd think after doing this for awhile I'd know which models effect what outcome. Surely the instruct model would be the smart one, ensuring the characters stay consistent along the way, and the only thing it wouldn't have going for it compared to the RP tailored models would be prose quality. It's probably fine leaving it at the top spot for percentages, yeah? After all, most Mixtral merges needed it's Instruct model at a decently high percentage in order to stand out, so surely it's the same with Nemo, right? ...RIGHT?! Wrong...
Turns out Magnum is decently smart on it's own, and I was actually lowering my merges' cohesion by insisting upon having Instruct - the smart one - and it's Bophades variant so high in the mix. Now, it still helped the model's IQ in other general areas, but for some stupid reason the character cards took a hit. Much like with Lyra, which could effect the model in ways I'm not entirely keen on (long rambling and Novel/RP format switching mid chat) lowering Instruct/Bophades to around .2 weight/.5 density seemed to be the maximum to where it had a positive effect on the model, and before these unwanted traits became noticeable. Mind you, I'm a real nit picker, so these issues probably wouldn't matter to other users just casually using the models, or even care to notice.
I was somewhat happy with the past Merges I did with Nemo, thinking these minor inconsistencies must just be a quirk because it's not in the 30B range. Nothing a swipe won't fix, I guess. Turns out I was just chasing after the wrong model for the top spot. Go figure... Even increasing Bophades and Lyra incrementally (.23/.56) showed slightly worse responses when the samples were turned off, so consider this merge to be a min-max version for now I guess, since I can't seem to find a better setup for it. Probably the last merge I'm doing for these 12B Mistral format models too, unless something crazy shows up using that format in the future.