Model Overview
This model, gemma-3-12b-it-vl-GLM-4.7-Flash-Polaris-Thinking-Heretic-Uncensored, is a 12 billion parameter Gemma-3 instruction-tuned variant developed by DavidAU. It features a 128k context window and is fine-tuned with a combination of the GLM 4.7 Flash reasoning dataset and the Polaris non-reasoning dataset. A key characteristic is its "deep thinking" capability, which is integrated into general operation, output generation, image processing, and benchmark performance.
Key Capabilities & Features
- Uncensored Output: Designed to provide direct responses without refusal, offering full freedom in content generation, though it may require explicit directives for highly graphic or explicit content.
- Enhanced Reasoning: Incorporates advanced reasoning logic, which can be explicitly activated with "think deeply: prompt" or through optional system prompts.
- Temperature Stability: Reasoning remains stable across a wide temperature range (0.1 to 2.5).
- Improved Benchmarks: Shows notable improvements over its Heretic uncensored base model in benchmarks like ARC Challenge, HellaSwag, and Winogrande.
- Low KL Divergence: Achieves a KL divergence of 0.0826, indicating minimal damage to the original model's distribution despite de-censoring.
Optimal Usage
For best performance, especially in chat and roleplay scenarios, users are advised to set a Smoothing_factor to 1.5 in applications like KoboldCpp, oobabooga/text-generation-webui, or Silly Tavern. Optional system prompts are provided to further enhance thinking and output generation, though automatic thinking activation is common due to fine-tuning.