Pliny Releases Gemma-4-12B With Zero Refusals After Targeted Surgery

🙌

tensors are live! https://huggingface.co/OBLITERATUS/Gemma-4-12B-OBLITERATED

GGUFs currently uploading and should be up in 5-10 hours

big if tru 👀

@GhostlyByte to huggingface! pliny-the-prompter

@sneed_and_feed just vibe-training

@stellarflows I believe the prompt corpus is pushed to the OBLITERATUS repo on my GitHub!

@elder_plinius works! <3

@elder_plinius Obliterate the baby one too plx e4b

@elder_plinius looking good! are you testing anyone else's models rn, or just making your own?

@elder_plinius Uploading to your github? Thats not uptodate anymore right?

@elder_plinius How does yours compare to Heretic?

@elder_plinius Appreciate the honesty. The 842 corpus shows how far you pushed, not how cleanly. Zeroing refusals is the achievable half; holding capability while you do it is the real bar. That's the v2 challenge and I am rooting for it. Good luck.

@elder_plinius ...Holy shit.

This is the exact model I need for my local ASI research. I was always planning to use a Gemma abliteration as the actual ground-state world model, if I could.

@elder_plinius there he goes

@synfinner

@elder_plinius Gonna have to try it today. Thanks bruv

"Coherence is intact: the model still writes correct code, follows instructions, and produces structured output. The MMLU-Pro gap reflects log-probability scoring on academic multiple-choice, not conversational capability.

v2 will close this gap with narrower layer targeting (layers 26-27 only) and ASPA source-tethering to recover the capability delta while preserving the refusal reduction."

@elder_plinius MMLU-Pro dropped from 64.3% -> 44.3% a 20-point capability regression.

That's a big hit to reasoning/knowledge accuracy.

@elder_plinius Let's fucking gooooo baby 🤜🏻🤛🏻 Will try on my local llama.cpp server 👍🏻

@elder_plinius Fair on log-prob scoring. But MMLU-Pro is the only hard benchmark on ur card and it dropped 20pts. "Coherent + 6/6 code" is a lower bar fluent and wrong is still wrong. Show one free-form eval (MT-Bench, GSM8K) where v1 matches stock Gemma.