grimjim/gemma-3-12b-it-biprojected-abliterated

Overview

This model, grimjim/gemma-3-12b-it-biprojected-abliterated, is a specialized variant of Google's gemma-3-12b-it.

Key Modifications

The primary differentiator of this model is the application of a technique called "projected abliteration." This method was used to:

• Reduce Refusal Rates: The model has been engineered to refuse user prompts far less often compared to its base model.
• Maintain Safety Awareness: Despite the reduced refusal, the model is designed to retain an understanding of safety and potential harms, aiming for a balance between permissiveness and responsible AI behavior.
• Targeted Intervention: The process involved determining refusal direction and removing projected contributions onto the harmless direction of specific layers, minimizing overall model damage.

No subsequent fine-tuning was applied after this modification process.

Use Cases

This model is particularly well-suited for applications where:

• Users require a more direct and less restrictive response style from the AI.
• The goal is to minimize instances of the model refusing to answer, while still acknowledging safety guidelines.
• Developers need a Gemma-based model with a modified refusal behavior for specific interactive or creative tasks.