cyrilgiordano83/gemma-4-E2B-it-uncensored
The cyrilgiordano83/gemma-4-E2B-it-uncensored model is a 5.1 billion parameter Gemma-4-E2B-it variant, fine-tuned to significantly reduce refusal behavior while preserving response quality. It achieves this through a norm-preserving biprojected obliteration method, making it suitable for applications requiring less restrictive content generation. This model is primarily designed for use cases where a more open and less censored language model is desired, offering a 32768 token context length.
Loading preview...
Overview
This model, cyrilgiordano83/gemma-4-E2B-it-uncensored, is a modified version of Google's Gemma-4-E2B-it, specifically engineered to remove refusal behaviors. It utilizes a novel "norm-preserving biprojected obliteration" method, which ensures that the model's weight magnitudes are preserved during the uncensoring process, preventing degradation in response quality.
Key Capabilities
- Significantly Reduced Refusals: Achieves a refusal rate of 1/100 on mlabonne's 100 prompts and 3/686 (0.4%) across a diverse cross-dataset validation, compared to 98/100 for the original model.
- Quality Preservation: Maintains response quality, with a harmless response length ratio of approximately 1.01, indicating no degradation in output length or coherence.
- Advanced Uncensoring Method: Employs a sophisticated technique that projects out refusal directions from weight components while preserving overall weight magnitudes, differing from standard projection methods.
- Efficient Processing: Uses per-layer refusal directions and a deterministic single-pass approach, making the uncensoring process faster and more effective than some alternative methods.
Good For
- Applications requiring less restrictive content generation: Ideal for scenarios where the default refusal mechanisms of base models are undesirable.
- Research into model safety and alignment: Provides a clear example of targeted refusal removal for studying model behavior.
- Creative and open-ended text generation: Suitable for tasks that benefit from a model less prone to self-censorship.