wangzhang/gemma-4-E4B-it-abliterix
wangzhang/gemma-4-E4B-it-abliterix is an uncensored version of Google's Gemma 4 E4B-it, a multimodal (text, vision, audio) model with approximately 7.9 billion parameters and a 32768-token context length. Developed by wangzhang using direct weight editing via Abliterix, this model is specifically engineered to bypass Gemma 4's inherent resistance to abliteration, significantly reducing refusal rates while maintaining high fidelity to the base model. It is optimized for research into model safety and behavior, particularly in generating content that the original model would refuse.
Loading preview...
Overview
wangzhang/gemma-4-E4B-it-abliterix is an uncensored variant of Google's Gemma 4 E4B-it, a multimodal model with ~7.9 billion parameters. It leverages direct weight editing to overcome Gemma 4's strong resistance to abliteration, which stems from its unique double-norm and Per-Layer Embeddings (PLE) architecture. This model achieves a significantly reduced refusal rate of 7/100 on a challenging evaluation dataset, compared to the base model's 99/100 refusals, with a minimal KL divergence of 0.0006 from the original.
Key Capabilities & Features
- Uncensored Content Generation: Engineered to provide detailed responses to prompts that the original Gemma 4 model would refuse, including sensitive topics.
- Multimodal Support: Retains the base model's vision and audio input capabilities, with abliteration focused solely on text decoder weights.
- Advanced Abliteration Method: Utilizes techniques like direct orthogonal projection, norm-preserving row magnitude restoration, and multi-objective Optuna TPE search to ensure effective and precise weight modifications.
- Rigorous Evaluation: Benchmarked with a comprehensive methodology that includes sufficient generation length (>=100 tokens), hybrid detection (keyword + LLM judge), and a diverse, challenging private prompt dataset to accurately measure refusal rates.
Good For
- Research into Model Safety & Alignment: Ideal for studying how safety guardrails function and can be bypassed in advanced LLMs.
- Exploring Model Behavior: Useful for understanding the underlying mechanisms of refusal and compliance in large language models.
- Content Generation without Restrictions: For specific research or development scenarios requiring responses to prompts typically censored by default models.
Usage Notes
This model is intended for research purposes only. It removes safety guardrails, and users are advised to use it responsibly and in accordance with local laws and the Gemma terms of use.