Model Overview
DavidAU/Gemma-3-27b-it-Uncensored-HERETIC-Gemini-Deep-Reasoning is a 27 billion parameter Gemma-3 instruction-tuned model, fine-tuned by DavidAU using Unsloth. This model is notable for being fully uncensored and maintaining intact image processing capabilities. It boasts a 128k context window and offers a wide temperature range from 0.1 to 2.5, with reasoning remaining stable across this range.
Key Capabilities & Features
- Deep Reasoning: Enhanced reasoning affects both image "intelligence" and output generation. Users can activate deeper thinking with "think deeply: prompt" or through optional system prompts and JINJA templates.
- Uncensored Output: The model is designed to not refuse requests, including those for sensitive content. It may require specific directives (e.g., using explicit slang) to generate content at expected graphic or explicit levels.
- Image Processing: Image processing functionality is fully operational.
- Flexible System Prompts: Supports custom system prompts to enhance thinking and output generation, with examples provided for both general and character-specific (e.g., Joker) thinking.
Performance & Benchmarks
The model demonstrates improved performance over its uncensored base, as shown in benchmarks for tasks like arc_challenge, hellaswag, and winogrande. Notably, its KL divergence is 0.07, indicating minimal damage to the model during de-censoring, and it significantly reduces refusals from 98/100 to 9/100 compared to the original google/gemma-3-27b-it.
Optimal Operation
For best results, users are advised to set a "Smoothing_factor" to 1.5 in interfaces like KoboldCpp, oobabooga/text-generation-webui, or Silly Tavern. Increasing repetition penalty to 1.1-1.15 is an alternative if smoothing is not available. Detailed guidance on parameters and samplers for maximizing performance is available via external links provided in the original README.