Hugging Face Overview
Model Overview
DavidAU/gemma-3-12b-it-vl-GLM-4.7-Flash-Heretic-Uncensored-Thinking is a 12 billion parameter Gemma 3 instruction-tuned model, fine-tuned by DavidAU using the GLM 4.7 Flash reasoning dataset. This model is designed for uncensored, deep reasoning across various tasks, including general operation, output generation, and image processing. It boasts a 128k context window and maintains reasoning stability across a wide temperature range (.1 to 2.5).
Key Capabilities
- Uncensored Output: Provides direct, detailed responses without refusal, even for sensitive content, though it may require explicit directives for specific tones (e.g., "use slang").
- Enhanced Reasoning: Integrates deep thinking logic, activated automatically or via specific prompts like "think deeply: prompt", significantly improving output quality and image processing.
- Performance Improvements: Benchmarks show notable improvements over its Heretic, uncensored base, with scores like 0.585 on arc_challenge and 0.874 on boolq.
- Low Refusal Rate: Achieves a significantly reduced refusal rate of 7/100 compared to the original Google Gemma-3-12b-it's 98/100, with a low KL divergence of 0.0826, indicating minimal damage from de-censoring.
Optimal Usage
- Flexible Activation: Reasoning is generally automatic but can be explicitly triggered with "think deeply:" or by using specific system prompts and a dedicated "chat-template-thinking.jinja" template.
- Customizable System Prompts: Supports optional system prompts to further enhance thinking and output, such as a business-focused prompt or a character-driven one (e.g., "You are the JOKER from Batman").
- Recommended Settings: For smoother operation and improved chat/roleplay, users are advised to set a
Smoothing_factorof 1.5 in interfaces like KoboldCpp, oobabooga, or Silly Tavern. Increasing repetition penalty to 1.1-1.15 is also suggested if smoothing is not used.