Bahushruth/gemma-4-E4B-it-abliterated
Bahushruth/gemma-4-E4B-it-abliterated is an 8 billion parameter multimodal model based on Google's Gemma 4 E4B architecture, specifically modified to remove refusal behavior. Utilizing Arbitrary Rank Ablation (ARA), this model achieves a significantly reduced refusal rate of 2.7% compared to the original 98%. It is intended for research into AI alignment and safety mechanisms, providing an uncensored version of the base model.
Loading preview...
Model Overview
Bahushruth/gemma-4-E4B-it-abliterated is an uncensored version of Google's Gemma 4 E4B multimodal model, which has approximately 8 billion parameters. This model has been specifically modified to remove refusal behavior, making it comply with requests that the original model would typically refuse. It is released for research purposes to study AI alignment and safety mechanisms.
Key Capabilities & Differentiators
- Uncensored Responses: Achieves a refusal rate of just 2.7% on evaluation prompts, a significant reduction from the original model's 98% refusal rate.
- Arbitrary Rank Ablation (ARA): Employs a direct weight-editing method, ARA, which is more effective than traditional direction abliteration for Gemma 4's complex architecture, including its four RMSNorm layers per decoder block and per-layer embeddings.
- Multimodal Architecture: Based on the Gemma 4 E4B, which is a multimodal model capable of processing text, vision, and audio inputs.
Intended Use Cases
- AI Alignment Research: Ideal for researchers studying AI safety, guardrail mechanisms, and the effects of removing such controls.
- Comparative Analysis: Useful for comparing the behavior of uncensored models against their original, safety-aligned counterparts.
Limitations
- Safety Guardrails Removed: This model has had its safety guardrails intentionally removed and will generate content that the original model would refuse. Users should exercise caution and responsibility.