Bathory X 26B-A4B v0.2 experimental model
a dramatic improvement over the 0.1 version.
This model is an experiment with model surgery. A true Frankenstein of layer merging and direct editing.
The following models contributed body layers:
hiwaifu-research/WaifuGemma4-26b-a4b-v1 - special thanks to MrDevolver for the suggestion
ReadyArt/For-Her-Darkside-26B-A4B-v1.4
ReadyArt/Dark-Scarlett-v1.0-26B-A4B
This time the embeddings were untied and the lm_head was distilled via top-k logit training from:
Gryphe/Gemma-4-26B-A4B-StyleTune-V2
The following model's edits were combined via task arithmetic to supply the abliteration:
coder3101/gemma-4-26B-A4B-it-heretic
Notes:
I ran a second attempt through the same merges, this time incorporating WaifuGemma, and choosing instead of transplanting the lm_head from Gryphe's StyleTune to use logit distillation on prose and roleplay datasets. My foreign-language token suppression was done only on Orion specifically prior to incorporating it into the model. The results in my own personal testing have been a much more solid model with much better writing skills - avoiding an awful lot of cliches, similes, and the everpresent smell of ozone.