zaakirio/gemma-4-12b-it-uncensored
The zaakirio/gemma-4-12b-it-uncensored model is a 12 billion parameter instruction-tuned variant of Google's Gemma-4-12B-it, specifically modified to suppress refusal behaviors. Developed by zaakirio using the Heretic abliteration tool, this model maintains the original Gemma's general capabilities while significantly reducing content refusal. It is designed for research, red-teaming, and creative applications where compliance with a broader range of prompts is desired.
Loading preview...
Overview
This model, zaakirio/gemma-4-12b-it-uncensored, is a decensored version of Google's gemma-4-12B-it (12 billion parameters, 32768 tokens context length). It was created using the Heretic tool, an automated implementation of directional ablation (also known as "abliteration"). This technique surgically suppresses the model's refusal behavior by orthogonalizing its weight matrices against a specific direction in the residual stream, as described by Arditi et al. (2024).
Key Capabilities & Differentiators
- Reduced Refusal Behavior: Achieves a significantly lower refusal rate (23/100) compared to the original Gemma-4-12B-it (99/100), while preserving general capabilities.
- Minimal Distribution Shift: Abliteration ensures that the model's core performance remains largely intact, with a low KL divergence of 0.043 from the original.
- Automated Decensoring: Utilizes the Heretic tool for an automated, non-fine-tuning approach to remove safety guardrails.
Intended Use Cases
This model is suitable for legitimate research, red-teaming, evaluation, and creative work where a model with suppressed refusal behavior is required. Users should be aware that safety guardrails have been removed, and they are responsible for the content generated.