wangzhang/gemma-4-12B-it-abliterix

Hugging Face
TEXT GENERATIONPricing:Input $1.2 / Cached $0.24 / Output $4.8Concurrent Unit Cost:1Model Size:12BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 23, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Warm

wangzhang/gemma-4-12B-it-abliterix is a 12 billion parameter Gemma-4 instruction-tuned model, developed by Wangzhang Wu, that has been uncensored and refusal-suppressed using directional ablation with abliterix. It maintains the original model's capabilities while significantly reducing safety-related refusals, making it suitable for safety research, red-teaming, and evaluation. The model achieves a 73.7 percentage point reduction in refusals compared to the base model, with minimal perturbation to its original behavior.

Loading preview...

Overview

This model, wangzhang/gemma-4-12B-it-abliterix, is an uncensored and refusal-suppressed version of Google's gemma-4-12B-it model, created by Wangzhang Wu using the abliterix tool. It employs directional ablation to remove safety-refusal behavior by orthogonally projecting a single refusal direction out of specific write-path projections across the decoder stack. This method ensures that the model's core capabilities remain largely intact while significantly reducing its tendency to refuse prompts.

Key Capabilities

  • Reduced Refusals: Achieves a substantial reduction in refusal rates, dropping from 99/100 to 26/100 on a set of 100 harmful prompts, as judged by an LLM (gemini-3.1-flash-lite).
  • Capability Preservation: A norm-preserving transform keeps the model's behavior very close to the original, with a low first-token KL divergence of 0.0735.
  • Improved over Alternatives: Demonstrates roughly half the residual refusals compared to the reference Heretic abliteration (zaakirio/gemma-4-12b-it-uncensored) under identical evaluation conditions.
  • Optimized Operating Point: The model represents a carefully chosen balance on the Pareto front, offering strong refusal removal with modest capability perturbation.

Good for

  • Safety Research: Ideal for studying refusal mechanisms and model robustness.
  • Red-Teaming: Useful for evaluating model vulnerabilities and alignment.
  • Evaluation: Suitable for assessing LLM behavior in scenarios where safety guardrails have been modified or removed.