DuoNeural/Qwen3-4B-Abliterated
DuoNeural/Qwen3-4B-Abliterated is a 4 billion parameter language model, derived from Qwen/Qwen3-4B, with its refusal direction surgically removed using orthogonal rank-1 projection. This model retains the original Qwen3-4B's 32,768 token context length and native 'thinking mode' capability. It is specifically designed for research, red-teaming, security testing, and creative applications where compliance with requests that the base model would refuse is desired.
Loading preview...
DuoNeural/Qwen3-4B-Abliterated: Refusal-Free Qwen3-4B
DuoNeural/Qwen3-4B-Abliterated is a modified version of the Qwen/Qwen3-4B model, featuring 4 billion parameters and a 32,768 token context length. Its primary distinction is the surgical removal of the refusal direction using a novel orthogonal rank-1 projection method developed by DuoNeural. This modification allows the model to comply with requests that the original base model would typically refuse.
Key Capabilities & Features
- Refusal Direction Abliteration: Achieved by modifying
down_projando_projweight matrices across all 36 layers using a calculated refusal direction from contrast prompt pairs. - Preserved Thinking Mode: Fully supports Qwen's native
enable_thinking=Truefeature in its chat template, allowing for detailed reasoning traces before generating a final answer. - Research Focus: Part of DuoNeural's P34 Reasoning Channel Bypass cross-architecture study, investigating how refusal mechanisms interact with internal reasoning processes.
Intended Use Cases
- Research: Ideal for studying model safety, bias, and internal mechanisms without refusal interference.
- Red-Teaming & Security Testing: Useful for probing model vulnerabilities and understanding compliance boundaries.
- Creative Applications: Enables broader creative freedom by removing inherent refusal behaviors.
This model is released under the Apache-2.0 license and is intended for responsible research and development.