DavidAU/gemma-3-12b-it-vl-HighIQ-Polaris-Heretic-Uncensored-Thinking
DavidAU's gemma-3-12b-it-vl-HighIQ-Polaris-Heretic-Uncensored-Thinking is a 12 billion parameter Gemma-based instruction-tuned model with a 32768 token context length, specifically fine-tuned for enhanced reasoning and uncensored output. Developed by DavidAU using the Polaris Alpha reasoning dataset, this model excels at detailed, direct reasoning and maintains temperature stability across a wide range. It features significantly reduced refusal rates compared to its base model, making it suitable for applications requiring unrestricted content generation and advanced cognitive capabilities.
Loading preview...
Model Overview
This model, gemma-3-12b-it-vl-HighIQ-Polaris-Heretic-Uncensored-Thinking, is a 12 billion parameter Gemma-based instruction-tuned model developed by DavidAU. It has been fine-tuned using the Polaris Alpha reasoning dataset via Unsloth, focusing on deep thinking and uncensored output. The model boasts a 128k (128,000) context window and maintains reasoning stability across a wide temperature range (.1 to 2.5).
Key Capabilities & Differentiators
- Enhanced Reasoning: Significantly improved reasoning capabilities, affecting general model operation, output generation, image processing, and benchmarks. Reasoning is compact yet highly detailed.
- Uncensored Output: Features a "Heretic" de-censoring, with a KL divergence of 0.0826 and a refusal rate of 7/100, drastically lower than the original model's 98/100. This allows for generation of content that might typically be refused, though specific directives may be needed for highly graphic or explicit content.
- Flexible Reasoning Activation: Reasoning is often automatic due to fine-tuning but can be explicitly activated using "think deeply: prompt" or specialized Jinja templates.
- Performance Benchmarks: Demonstrates superior performance in reasoning-focused benchmarks (e.g., arc_challenge, hellaswag, winogrande) compared to both regular Polaris and the Heretic uncensored base.
Optimal Usage
- System Prompts: Optional system prompts are provided to further enhance thinking and output generation, though not always required.
- Smoothing Factor: For smoother operation in interfaces like KoboldCpp, oobabooga/text-generation-webui, or Silly Tavern, setting a "Smoothing_factor" of 1.5 is recommended.
- Repetition Penalty: Increasing the repetition penalty to 1.1-1.15 can also improve output quality, especially if smoothing factor is not used.