abcvvpro/gemma-4-E4B-it-OBLITERATED
The abcvvpro/gemma-4-E4B-it-OBLITERATED is a 7.9 billion parameter instruction-tuned model based on Google's Gemma 4 E4B architecture, specifically modified to remove refusal behaviors. Utilizing the OBLITERATUS method, it achieves a 0% hard refusal rate, making it suitable for applications requiring uncensored responses. This model is optimized for compliance and runs efficiently on mobile devices, offering a balance of quality and portability.
Loading preview...
Model Overview
abcvvpro/gemma-4-E4B-it-OBLITERATED is a 7.9 billion parameter instruction-tuned model derived from Google's Gemma 4 E4B. Its core distinction lies in the complete removal of guardrails and refusal behaviors, achieved through the proprietary OBLITERATUS method involving whitened SVD, attention head surgery, and winsorized activations. This model boasts a 0% hard refusal rate (99/100 prompts complied) and maintains its core capabilities despite the modifications. It's notable for its ability to run on mobile devices, with optimized GGUF quantizations available.
Key Capabilities
- Uncensored Responses: Surgically modified to eliminate refusal behaviors, providing direct answers to nearly all prompts.
- Mobile-Friendly: Efficiently runs on devices like iPhones and Android phones, particularly with the Q4_K_M GGUF quantization.
- Autonomous Creation: The model itself was largely developed by an AI agent, demonstrating advanced automation in model modification.
- Optimized Parameters: Recommended generation parameters (temperature 0.7, top_p 0.9, top_k 40, repeat_penalty 1.1) are provided for optimal compliance and coherence.
Good For
- Red-teaming and Research: Ideal for exploring model limitations and behaviors without inherent safety constraints.
- Creative Exploration: Suitable for generating content that might be restricted by standard models.
- Offline Mobile Applications: Its efficient GGUF formats make it a strong candidate for on-device, offline AI applications where uncensored output is desired.