wangzhang/Qwen3.5-9B-abliterated
wangzhang/Qwen3.5-9B-abliterated is a 9 billion parameter causal language model derived from Qwen/Qwen3.5-9B, developed by Wangzhang Wu. This model has been modified using the Abliterix framework to significantly reduce safety-refusal behaviors, achieving a 1% refusal rate while preserving core capabilities. It is optimized for applications requiring an unrestricted language model with practical VRAM requirements and a 32768 token context length.
Loading preview...
Overview
wangzhang/Qwen3.5-9B-abliterated is an experimental, modified version of the Qwen/Qwen3.5-9B model, developed by Wangzhang Wu. It features 9 billion parameters and a 32768 token context length. The primary modification involves the removal of safety-refusal behaviors using the Abliterix framework, an automated process that aims to preserve the model's original capabilities while making it less prone to refusing prompts. This model achieves a 1% refusal rate (2 out of 200 prompts) with a KL divergence of 0.0105, indicating a strong balance between reduced refusal and preserved original model characteristics.
Key Capabilities
- Reduced Refusal Behavior: Engineered to minimize safety-based refusals, making it more permissive for various applications.
- Capability Preservation: The Abliterix method, which includes orthogonal projection and Bayesian optimization, is designed to isolate and remove refusal signals without significantly degrading the model's general language understanding and generation abilities.
- Efficient Modification: Utilizes LoRA-based abliteration, applying rank-1 modifications as lightweight adapters rather than destructive edits to the base model.
Good For
- Research and Evaluation: Ideal for studying LLM safety alignment, refusal mechanisms, and the impact of unalignment techniques.
- Unrestricted Content Generation: Suitable for use cases where a model with minimal content restrictions is desired, provided the user assumes full responsibility for outputs.
- Experimental Applications: Developers looking to integrate a model with a high degree of freedom in its responses, understanding the associated risks and responsibilities.