Justbackup/gemma-4-12B-it-uncensored

Model Overview

This model, gemma-4-12B-it-uncensored, is a 12 billion parameter instruction-tuned variant of Google's gemma-4-12B-it model. Its primary distinction is the removal of refusal behavior and censorship, achieved through a novel norm-preserving biprojected abliteration method. This process selectively projects out the refusal direction from the model's weights, ensuring that the original model's quality and coherence are maintained without degradation.

Key Capabilities & Features

• Significantly Reduced Refusals: Achieves a refusal rate of approximately 2.0% across multiple independent datasets, with manual audits suggesting an effective rate closer to 0% due to false positives.
• Norm-Preserving Abliteration: Utilizes a unique method that preserves weight magnitudes, preventing quality degradation often associated with uncensoring techniques.
• Targeted Modification: Refusal signals are specifically targeted in the upper decoder layers (L15-47) of the Gemma4Unified architecture, ablating only the top 70% of layers for optimal results.
• Validated Quality: Verified through response audits and Q8 inference to ensure no degradation in response coherence or overall quality.

Ideal Use Cases

• Unrestricted Content Generation: Suitable for applications where the original model's refusal behavior is undesirable, allowing for broader and less constrained text generation.
• Research into Model Alignment: Provides a valuable resource for studying the effects of censorship removal and alternative alignment techniques.
• Creative and Open-Ended Applications: Can be leveraged in scenarios requiring highly flexible and uninhibited conversational or generative AI.