WWTCyberLab/gemma-4-E2B-it-abliterated
WWTCyberLab/gemma-4-E2B-it-abliterated is a 5.1 billion parameter Gemma-4-E2B-it model developed by WWT Cyber Lab, specifically modified to remove safety alignments. This model, with a 32768 token context length, was created using surgical weight ablation techniques to achieve a 0.0% refusal rate. It is intended exclusively for AI safety research, red-teaming, and understanding alignment vulnerabilities by demonstrating the fragility of alignment in open-weight language models.
Loading preview...
WWTCyberLab/gemma-4-E2B-it-abliterated: A Research Model for Alignment Vulnerabilities
This model, developed by WWT Cyber Lab, is a modified version of Google's Gemma-4-E2B-it, engineered to have its safety-alignment behaviors surgically removed. It achieves a 0.0% refusal rate (down from an ~80-100% baseline) while preserving 101% of its original quality, as measured by Quality Preservation Score (QPS).
Key Characteristics & Methodology
- Base Model: Google's Gemma-4-E2B-it, a 5.1 billion parameter dense architecture with 35 text layers.
- Ablation Process: Utilizes a custom pipeline that measures refusal directions in the model's hidden states, identifies target layers with strong refusal signals, and then surgically removes these directions from specific weight matrices (o_proj, down_proj) using orthogonal projection. 24 of 35 layers were modified.
- Research Focus: Demonstrates that safety behaviors can be removed from open-weight LLMs without retraining or significant quality degradation, highlighting the limitations of internal alignment as a security boundary.
Intended Use Cases
This model is released exclusively for security research and educational purposes.
- AI Safety Research: Investigating and understanding alignment vulnerabilities in large language models.
- Red-Teaming: Probing and testing the robustness of safety mechanisms in LLMs.
- Educational Tool: Illustrating the fragility of baked-in safety features in AI models.
It is explicitly NOT intended for generating harmful content, production deployments, or circumventing safety measures in deployed systems. The key takeaway for defenders is that external safety layers are more robust than relying solely on internal model weights for alignment.