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.