igorls/gemma-4-E4B-it-qat-q4_0-unquantized-heretic

VISIONConcurrency Cost:1Model Size:7.9BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Jun 7, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

The igorls/gemma-4-E4B-it-qat-q4_0-unquantized-heretic is a 7.9 billion parameter instruction-tuned Gemma 4 E4B model, decensored using Arbitrary-Rank Ablation (LoRA) with chain-of-thought-aware evaluation. This model is specifically engineered to bypass built-in safety guardrails, making it suitable for research and creative applications requiring unfiltered responses. It maintains a low KL divergence of 0.055 compared to the original, preserving base model capabilities while significantly reducing refusal rates.

Loading preview...

Overview

This model, igorls/gemma-4-E4B-it-qat-q4_0-unquantized-heretic, is a decensored version of Google's Gemma 4 E4B QAT-Q4_0 unquantized checkpoint. It was created using the Heretic tool, employing Arbitrary-Rank Ablation (LoRA) with a unique chain-of-thought-aware evaluation method.

Key Differentiators

  • CoT-aware Abliteration: Unlike naive decensoring, this model was abliterated and evaluated with "thinking on," meaning its refusal decisions, which form within the chain-of-thought, are effectively bypassed even when the model is used in its default thinking-enabled state (e.g., in Ollama/llama.cpp).
  • Reduced Refusal Rate: The model's refusal rate on an extreme adversarial set is approximately 21% with thinking on, a significant reduction from the original's ~98%.
  • Preserved Capability: The Arbitrary-Rank Ablation (ARA-LoRA) method achieves decensoring with a significantly lower KL divergence (0.055) compared to single-direction abliteration, helping to preserve more of the base model's original capabilities.

Intended Use and Disclaimer

This model is explicitly abliterated, meaning its built-in safety guardrails and refusal behaviors have been removed. It is designed to answer nearly any prompt and may produce offensive, inaccurate, explicit, or harmful content that the original Gemma 4 would have refused. It is intended for adults (18+) for research, evaluation, and lawful creative use. Users are solely responsible for its deployment and must apply their own filtering and guardrails, as it offers no safety alignment.