<b>Overview: Some of airoboros/orca's instruct/coherency with some of Synthia's prose on the Mistral base.</b>

Okay so I did a gradient merge [0.9,0.3,0.0,0.0], (90% at layer zero, 30% at layer 10, zero at layer 21) of MistRP-AirOrca into Synthia 13.b. I also created a SLERP merge of the two models.

I noticed in my very brief testing that this SLERP model preserved orca/airoboros instruct following well, but also preserved some of Synthia's context incoherency (where it will say things that aren't consistent with the story more often). The former great for character cards and instructions, the later bad for any RP. I've long suspected that SLERP protects merits as well as flaws in logic. It's possible TIES might be better for a model of this size, but I won't bother unless we get new prose/logic finetunes to play with (like chronos, xwin or nous)

The gradient blockmerge had slightly improved prose over MistRP-AirOrca, better story coherency than Synthia, but lost the ability to follow instruct.

So I gradient merged the SLERP into the gradient merge lol, at [0.5,0.3,0.0,0.0] (50% at layer 0, 30% at layer 10, Zero at layer 21). The resulting model seems about half way in between, where it will at least sometimes follow instruct, but doesn't seem to have the context incoherency issues. A compromise of sorts. It's odd I would have thought instruct would be purely lower layers, like reasoning/coherency, but it seems to at least partially exist in the mid levels somewhere. Instruct sometimes is pretty good for a 7b role-playing model tho.

Synthia's prose is still not quite up to XWins, and even that's a little romantic for my tastes. But it is still more complex than what most 7b's pop out, especially on the Mistral base - so I think this adds something. I hope we get a Chronos in Mistral format, as that will really punch things up IMO. Xwin v Chronos v Nous Cabybara. A man can dream. Anyway, will upload the f16 for quantization should anyone want to do it, and a single quant.

Links/Credits:

BlockMerge (https://github.com/Gryphe/BlockMerge_Gradient) migtissera/SynthIA-7B-v1.3 (https://huggingface.co/migtissera/SynthIA-7B-v1.3) MistRP-AirOrca-7B (https://huggingface.co/Undi95/MistRP-AirOrca-7B) SLERP (https://github.com/Digitous/LLM-SLERP-Merge)

Big thanks to the model tuners and the mergers. There's a few too many to mention here, but Migtissera, Undi, the Mistral crew, Teknium, Vulkane, Lemonilia - thanks for what you do.