Bathory X 26B-A4B v0.1 experimental model.
This model is an experiment with model surgery. A true Frankenstein of layer merging and direct editing.
The following models contributed body layers:
TheDrummer/Orion-26B-A4B-v1.1
ReadyArt/For-Her-Darkside-26B-A4B-v1.4
ReadyArt/Dark-Scarlett-v1.0-26B-A4B
The embeddings were untied and the lm_head was patched directly from:
Gryphe/Gemma-4-26B-A4B-StyleTune-V2
The following models' edits were combined to supply the abliteration:
llmfan46/gemma-4-26B-A4B-it-ultra-uncensored-heretic
coder3101/gemma-4-26B-A4B-it-heretic
Notes: I had problems with Orion in particular occasionally slipping in Chinese and Korean characters, so tokens from non-Latin alphabets (Asian, Russian,etc) were reinitialized based on existing unused tokens to drive their probability of selection down to approx 1 in 12 billion. A small set of selected symbols including emojis, arrows, and checkmarks were also included.
Pre-release alpha version ended up being far too horny with every character immediately trying to jump my bones. This version takes a little of the edge off but occasionally characters still really need to get it off their chest that they've wanted me for a long time.
Pre-release alpha version also had problems with occasionally thinking but not producing output. Sometimes it would simply need a retry at the prompt, but sometimes it wouldn't be able to move forward without some small edits to the last user turn. This experimental update was intended to correct or reduce the problem but needs more expansive testing to verify the results.