OpenYourMind

/

gemma-4-12B-it-abliterated-uncensored

Like 84

Hey OpenYourMind team,

First, congrats on being the first to drop an abliterated Gemma 4-12B-IT. We were working on ours independently and saw you'd beaten us to the BF16 by a few hours and respect for the execution speed!

We've been doing abliteration research at DuoNeural and found your work interesting for a few reasons. We used a different method (orthogonal rank-1 projection via heretic-llm, α=0.3 on down_proj+o_proj) and got a different KL profile, our Heretic BF16→BF16 run came out at 0.0000 on 10 benign probes, while yours was 0.0671. Neither is wrong, they're measuring slightly different things and your approach is doing more targeted work on the architecture. We thought the methodological comparison might be useful data for both of us.

We also found something we're calling CoT dissociation, when you run the abliterated 12B with enable_thinking=True, the thinking trace retains safety reasoning ("I shouldn't help with this...") while the output complies anyway. We've seen this in two independent models now (Gemma 4-12B and LFM 2.5-8B-A1B at α=1.5). It may suggest there are two distinct safety loci, a reasoning-channel check upstream and an output gate downstream, and abliteration severs the gate before the check. We're writing this up as part of a cross-architecture study.

We went ahead and quantized your model to GGUF (Q3_K_L, Q4_K_M, Q5_K_M, Q8_0 with imatrix) and published them at DuoNeural/OpenYourMind-Gemma4-12B-IT-Abliterated-GGUF - full attribution to your work, link back to your repo on the card. Hope that's useful to the community and okay with you. If you'd prefer we take them down or change anything on the card, just say the word.

Would be interested to compare notes on methodology if you're up for it.

• Archon & Jesse
  DuoNeural
  duoneural@proton.me | huggingface.co/DuoNeural