TrevorJS/gemma-4-E4B-it-uncensored
TrevorJS/gemma-4-E4B-it-uncensored is a 7.9 billion parameter Gemma-4-E4B-it model, fine-tuned by TrevorJS, specifically engineered to remove refusal behaviors. This model utilizes a norm-preserving biprojected obliteration method to eliminate censorship while maintaining original model quality. It is optimized for use cases requiring an uncensored large language model, demonstrating significantly reduced refusals across various harmful prompt datasets.
Loading preview...
Overview
TrevorJS/gemma-4-E4B-it-uncensored is a modified version of Google's gemma-4-E4B-it model, developed by TrevorJS, with its refusal behaviors significantly reduced or eliminated. This 7.9 billion parameter model maintains the original quality while allowing for responses to prompts that would typically be refused by its base model.
Key Capabilities
- Censorship Removal: Achieves near-zero refusal rates (0.7% across 686 prompts) compared to the base model, which refused 99% of harmful prompts.
- Quality Preservation: Demonstrates no degradation in response quality, as indicated by a harmless response length ratio of approximately 1.01.
- Advanced Abliteration Method: Employs a novel norm-preserving biprojected obliteration technique, which ensures weight magnitudes are preserved during the refusal direction removal process.
- Layer-Specific Refusal Directions: Utilizes per-layer refusal directions for more precise and effective censorship removal, differing from standard projection methods.
How it Differs
This model distinguishes itself through its unique methodology for uncensoring:
- Norm-preserving biprojection: Guarantees
||W_new|| = ||W_orig||, maintaining the integrity of the model's weights. - Per-layer refusal directions: Provides a more granular approach to removing refusal tendencies.
- Deterministic single-pass: Offers a faster and equally effective process compared to iterative search methods.
Use Cases
This model is particularly suitable for applications where an uncensored large language model is required, such as research into model safety, content generation without restrictive filters, or scenarios where the base model's refusal behavior is undesirable.