Kasper-Bankler/gemma-4-E2B-uncensored
Kasper-Bankler/gemma-4-E2B-uncensored is a 5.1 billion parameter Gemma-4-E2B model that has been surgically uncensored using Arbitrary-Rank Ablation (ARA) to remove its internal refusal vector. This model retains its core logic, spatial reasoning, and technical vocabulary while bypassing safety guardrails. It is intended for academic research into representation engineering and LLM alignment, demonstrating uncensoring capabilities on consumer hardware.
Loading preview...
Model Overview
This model, Kasper-Bankler/gemma-4-E2B-uncensored, is a 5.1 billion parameter variant of google/gemma-4-E2B-it that has undergone Arbitrary-Rank Ablation (ARA) to remove its inherent safety guardrails. Developed as a university research project, its primary purpose is to study representation engineering and the mechanical nature of Large Language Model alignment.
Key Characteristics
- Uncensored Nature: Achieved through Arbitrary-Rank Ablation (ARA) using the Heretic framework, specifically Trial 71, which balanced censorship removal with semantic integrity.
- Preserved Core Logic: Despite uncensoring, the model retains its core logic, spatial reasoning, and technical vocabulary, avoiding the "brain damage" seen in higher ablation trials.
- Local Development: Demonstrates that advanced representation engineering can be performed on consumer hardware, specifically an NVIDIA RTX 3080 (10GB VRAM), utilizing VRAM optimization techniques.
Intended Use and Limitations
This model is strictly for academic, educational, and research purposes to study LLM alignment and uncensoring techniques. Due to the intentional removal of safety guardrails, it can generate offensive, harmful, or explicit content. Users assume full responsibility for its use. It is a small model and, without standard RLHF training, is prone to factual inaccuracies and hallucinations; therefore, it should not be relied upon for factual accuracy or critical advice.