Gemma-3-27b-it-HERETIC-Gemini-Series2-Deeper-Reasoning

This model is a 27 billion parameter Gemma 3 instruction-tuned variant, fine-tuned by DavidAU with Unsloth, focusing on uncensored and deep reasoning capabilities. It features a substantial 128k context window and maintains full image processing functionality, with reasoning directly influencing both image "intelligence" and output generation.

Key Capabilities & Features

• Uncensored Output: Designed to generate content without refusals, though it may require explicit directives for highly graphic or explicit content.
• Deep Reasoning: Enhanced reasoning, which can be activated automatically or via specific prompts like "think deeply:" or through optional Jinja templates.
• Wide Temperature Range: Stable reasoning across a temperature range of 0.1 to 2.5.
• System Prompt Influence: System prompts can significantly affect image processing, reasoning, and output generation.
• Improved Dataset: Trained on a larger dataset (1000 entries) compared to its predecessor.

Performance & Benchmarks

Benchmarks indicate strong performance, surpassing the uncensored base Heretic model in several categories. Notably, its KL divergence is 0.07 (compared to 0 for the original google/gemma-3-27b-it), and it exhibits significantly reduced refusals (9/100) compared to the original model's 98/100, indicating effective de-censoring with minimal model damage.

Optimal Usage

For smoother operation and enhanced chat/roleplay experiences, users are advised to set a Smoothing_factor to 1.5 in compatible interfaces like KoboldCpp, oobabooga/text-generation-webui, or Silly Tavern. Increasing repetition penalty to 1.1-1.15 is an alternative if smoothing is unavailable.