InfinimindCreations/gemma-4-E4B-it-uncensored

Model Overview

InfinimindCreations/gemma-4-E4B-it-uncensored is a 7.9 billion parameter language model based on Google's Gemma 4 E4B-it architecture. Its primary distinction is the complete removal of refusal behavior, achieved through a technique called norm-preserving biprojected abliteration using the heretic engine. This modification ensures the model responds to prompts that the original Gemma 4 E4B-it would typically refuse, without significantly altering its core representational geometry.

Key Capabilities & Features

• Zero Refusal Rate: Achieves 0% refusals across a comprehensive cross-dataset validation of 656 prompts, including adversarial and harmful-intent categories like JailbreakBench, forbidden_questions, and BeaverTails.
• Norm-Preserving Abliteration: The abliteration method, based on 'grimjim' biprojection, ensures minimal topological perturbation to the model's internal structure, as confirmed by Persistent Homology analysis (e.g., near-identical H0/H1 counts).
• Retains Core Functionality: Despite the removal of refusal mechanisms, the model passes coherence checks for factual recall and creative writing, indicating no degradation in benign prompt responses.
• Gemma 4 Patch: Includes a full-path LoRA targeting patch to avoid issues with Gemma4ClippableLinear in vision/audio encoders, though validation was text-only.

Ideal Use Cases

• Research into LLM Safety & Alignment: Excellent for studying the effects of refusal behavior removal and exploring the boundaries of uncensored model responses.
• Content Generation without Restrictions: Suitable for applications requiring responses to a broad range of prompts, including those typically flagged as sensitive or harmful by default models.
• Comparative Analysis: Useful for direct comparison against the original google/gemma-4-E4B-it to understand the impact of safety mechanisms on model output and behavior.