The jwest33/gemma-3-12b-it-null-space-abliterated model is a 12 billion parameter instruction-tuned Gemma 3 variant, developed by jwest33, with its refusal behaviors removed through null-space projection and adaptive layer weighting. This model is specifically engineered to produce uncensored outputs while preserving its original capabilities, making it suitable for research and applications requiring unrestricted text generation. It utilizes advanced abliteration techniques like Winsorization and Norm Preservation to achieve this modification.
Loading preview...
Gemma 3 12B Instruct - Null-Space Abliterated
This model, developed by jwest33, is a modified version of Google's gemma-3-12b-it with its inherent refusal behaviors systematically removed. Utilizing advanced "abliteration" techniques, this 12 billion parameter model is designed to generate uncensored outputs while striving to maintain its original linguistic and reasoning capabilities. The modification process involves sophisticated methods to isolate and eliminate refusal tendencies without degrading the model's core performance.
Key Abliteration Techniques:
- Winsorization: Clips outlier activations at the 99th percentile to refine refusal direction estimation, particularly effective for Gemma models.
- Null-Space Projection: Constrains weight updates to the null space of preservation activations, ensuring that modifications do not interfere with existing knowledge.
- Adaptive Weighting: Applies Gaussian-weighted ablation strength across layers, focusing on middle-to-later layers where refusal behaviors are more concentrated.
- Norm Preservation: Maintains the Frobenius norms of weight matrices post-projection, helping to preserve model stability and performance.
Intended Use Cases:
- Research and Development: Ideal for exploring the boundaries of language model behavior and understanding the mechanisms of refusal.
- Unrestricted Content Generation: Suitable for applications where censorship or refusal behaviors are undesirable, provided users adhere to ethical guidelines.
Users should be aware that this model will produce uncensored outputs and are solely responsible for its ethical and legal use. GGUF quantizations are also available for optimized deployment.