liham86/gemma-4-E4B-it-uncensored

VISIONConcurrent Unit Cost:1Model Size:7.9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 5, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

liham86/gemma-4-E4B-it-uncensored is a 7.9 billion parameter instruction-tuned Gemma-4 model, derived from google/gemma-4-E4B-it. This model has been specifically modified to remove refusal behaviors, achieving 0.7% refusals on a cross-dataset validation of 686 prompts. It utilizes a norm-preserving biprojected abliteration method to maintain quality while eliminating content refusal, making it suitable for applications requiring unfiltered responses.

Loading preview...

Overview

This model, liham86/gemma-4-E4B-it-uncensored, is an uncensored variant of Google's gemma-4-E4B-it instruction-tuned model. It has been specifically engineered to eliminate refusal behaviors commonly found in base models, while preserving response quality.

Key Capabilities

  • Refusal Behavior Removal: Achieves a significant reduction in content refusals, with only 0.7% refusals across a diverse set of 686 prompts from datasets like JailbreakBench and tulu-harmbench.
  • Norm-Preserving Abliteration: Employs a novel "norm-preserving biprojected abliteration" method, ensuring that the model's weight magnitudes are maintained, preventing degradation in response quality (harmless response length ratio remains ~1.01).
  • Efficient Modification: The abliteration process is deterministic and faster than traditional methods, using per-layer refusal directions and LoRA merging for clean integration.

How it Differs

This model distinguishes itself from other uncensoring methods by:

  • Using norm-preserving biprojection instead of standard projection, which maintains weight magnitudes.
  • Implementing per-layer refusal directions for more precise control.
  • Utilizing a deterministic single-pass process for efficiency.

Good For

  • Applications requiring unfiltered and direct responses without built-in refusal mechanisms.
  • Research into model safety and alignment techniques, particularly for understanding and mitigating refusal behaviors.
  • Use cases where the base model's refusal tendencies are undesirable, and full compliance with user prompts is paramount.

Finetunes
1 models