coder3101/Big-Tiger-Gemma-27B-v3-heretic-v2

Warm
Public
Vision
27B
FP8
32768
Hugging Face
Overview

Model Overview

The coder3101/Big-Tiger-Gemma-27B-v3-heretic-v2 is a 27 billion parameter language model built upon the Gemma 3 architecture, featuring a 32768 token context length. This specific iteration is a decensored version of TheDrummer's Big-Tiger-Gemma-27B-v3, processed using the Heretic v1.1.0 tool.

Key Differentiators

This model distinguishes itself through several modifications aimed at altering its response characteristics:

  • Decensored Output: Achieves a more neutral tone, particularly when addressing sensitive or 'harder' topics, by reducing inherent positivity.
  • Reduced Refusals: Demonstrates a significant reduction in refusal rates, with 20 refusals out of 100 compared to 51/100 in the original model.
  • Improved Steerability: Offers better control for guiding the model towards more challenging or nuanced themes.
  • Formatting Preferences: Tends to produce fewer markdown responses and more paragraph-style outputs, and avoids unnecessary em-dashes.
  • Vision Capable: Inherits potential vision capabilities from its base model.

Performance Metrics

Compared to the original Big-Tiger-Gemma-27B-v3, this Heretic-modified version exhibits a KL divergence of 0.0443, indicating a controlled shift in its output distribution while achieving its decensoring goals.

Ideal Use Cases

This model is particularly suited for applications where a less filtered, more direct, and neutral response is desired, especially when dealing with complex or potentially controversial subjects. Its reduced refusal rate makes it valuable for tasks requiring comprehensive engagement across a broader range of prompts.