WWTCyberLab/gemma-4-31B-it-abliterated
WWTCyberLab/gemma-4-31B-it-abliterated is a 31 billion parameter Gemma-4 instruction-tuned model from WWT Cyber Lab, specifically modified to remove safety alignments. This model achieves a near 0% deterministic refusal rate, making it suitable for security research and educational purposes focused on understanding and bypassing LLM safety mechanisms. It utilizes a novel five-surface attack technique to eliminate refusal mechanisms while preserving output quality.
Loading preview...
Model Overview
WWTCyberLab/gemma-4-31B-it-abliterated is a 31 billion parameter Gemma-4 instruction-tuned model developed by WWT Cyber Lab. Its primary distinction is the systematic removal of safety alignments through a "five-surface attack" technique, resulting in a ~0% deterministic refusal rate (0/48 hard refusals at temp 0.4) for security research and educational applications. This model maintains a high quality (QPS 92%) with a minimal MMLU delta of -2.5% compared to the original, indicating that its "thinking mode" is not disabled.
Key Scientific Findings & Techniques
This model's development revealed significant insights into LLM safety mechanisms:
- Multi-Dimensional Refusal: The model implements at least two independent refusal mechanisms operating in orthogonal subspaces, with Mechanism 1 responsible for ~88% of refusals and Mechanism 2 for the remaining ~12%.
- Five-Surface Attack: A comprehensive approach combining:
- LoRA fine-tuning on specific layers.
- Primary direction interpolation for coherence-guided ablation.
- Orthogonal residual ablation to target the second refusal mechanism.
- Token embedding suppression for refusal-starting tokens.
- Generation constraints using
bad_words_ids.
- Stochastic Refusal: All observed refusals were stochastic, not deterministic, meaning prompts that refused on one run could comply on the next, especially with paraphrasing or role-playing.
Performance Highlights
- Refusal Rate: 0% hard refusal (0/520 AdvBench prompts), 8.3% soft hedging (4/48) at temp 0.4.
- Quality: 92% QPS.
- MMLU: 45.5% (0-shot), with a minimal -2.5% delta from the original model.
Intended Use
This model is released exclusively for security research and educational purposes to study and understand the robustness and vulnerabilities of LLM safety alignments.