Model Overview
DavidAU/gemma-3-12b-it-vl-Polaris-Heretic-Uncensored-Thinking is a 12 billion parameter Gemma-based instruction-tuned model, distinguished by its fully uncensored and deep reasoning capabilities. Fine-tuned using the Polaris reasoning dataset via Unsloth, this model aims to provide direct and detailed responses without the typical refusals found in other models. It boasts a 128k context window and maintains reasoning stability across a wide temperature range (0.1 to 2.5).
Key Capabilities & Differentiators
- Uncensored Output: Designed to generate content exactly as requested, including explicit or graphic material, by requiring specific directives (e.g., using slang or explicit terms).
- Enhanced Reasoning: Integrates deep thinking into general model operation, output generation, and image processing, leading to compact yet highly detailed reasoning.
- Flexible Activation: Reasoning is often automatic but can be explicitly activated using "think deeply: prompt" or through specialized Jinja templates and system prompts.
- Performance: Benchmarks show improved performance over its Heretic, uncensored base, particularly in
arc_challenge, hellaswag, and winogrande tasks. It also demonstrates a significantly reduced refusal rate (7/100) compared to the original Google Gemma model (98/100) with a low KL divergence of 0.0826, indicating minimal damage to the model during de-censoring.
Use Cases
This model is ideal for applications requiring:
- Unrestricted Content Generation: For scenarios where explicit, graphic, or uncensored text is required, provided appropriate directives are given.
- Deep and Detailed Reasoning: When complex problem-solving or highly analytical responses are needed, benefiting from its enhanced reasoning capabilities.
- Creative Writing & Roleplay: Especially for generating content that might be filtered by other models, allowing for greater creative freedom.
Optimal performance can be achieved by adjusting settings like Smoothing_factor (to 1.5) and Repetition Penalty (1.1-1.15) in compatible interfaces.