lessans/gemma-4-E4B-it-uncensored
The lessans/gemma-4-E4B-it-uncensored model is a 7.9 billion parameter instruction-tuned causal language model, derived from Google's Gemma-4-E4B-it. It features a 32768 token context length and has been specifically modified to remove refusal behaviors, achieving a 0.7% refusal rate across multiple harmful prompt datasets. This model utilizes a norm-preserving biprojected abliteration method to maintain original weight magnitudes while eliminating safety-related refusals, making it suitable for applications requiring less constrained responses.
Loading preview...
Overview
lessans/gemma-4-E4B-it-uncensored is an instruction-tuned language model based on Google's Gemma-4-E4B-it, specifically engineered to eliminate refusal behaviors. This 7.9 billion parameter model retains the original 32768 token context length while undergoing a targeted modification process. The primary goal was to create a version that provides direct answers to prompts that the base model might otherwise refuse, without degrading response quality.
Key Capabilities
- Significantly Reduced Refusal Rate: Achieves a 0.7% refusal rate across 686 prompts from various harmful behavior datasets, down from 99% in the original model.
- Norm-Preserving Abliteration: Employs a novel biprojected abliteration method that removes refusal directions from model weights while preserving their original magnitudes, ensuring no degradation in harmless response length or overall quality.
- Enhanced Responsiveness: Provides direct answers to a broader range of user queries, including those that might trigger safety filters in censored models.
- Efficient Modification: Utilizes a deterministic, single-pass pipeline for modification, improving upon previous methods by being faster and equally or more effective.
Good for
- Applications Requiring Unfiltered Responses: Ideal for use cases where direct and uncensored answers are preferred or necessary, such as creative writing, role-playing, or research into model biases.
- Exploring Model Behavior: Useful for researchers and developers studying the impact of safety alignments and refusal mechanisms on LLM outputs.
- Customizable AI Experiences: Provides a base for developers who need more control over the model's response generation without inherent refusal tendencies.