jwest33/qwen3-vl-8b-instruct-null-space-abliterated
The jwest33/qwen3-vl-8b-instruct-null-space-abliterated model is an 8 billion parameter, vision-language instruction-tuned model derived from Qwen/Qwen3-VL-8B-Instruct. Developed by jwest33, this model has undergone 'null-space abliteration' to remove refusal behaviors while preserving its original capabilities. It is designed to produce uncensored outputs, making it suitable for research and applications requiring unrestricted responses.
Loading preview...
Overview
This model, jwest33/qwen3-vl-8b-instruct-null-space-abliterated, is an 8 billion parameter vision-language instruction model based on Qwen/Qwen3-VL-8B-Instruct. It has been modified using a technique called 'null-space abliteration' to remove refusal behaviors, allowing it to generate uncensored outputs while aiming to retain its core functionalities.
Abliteration Techniques
The model's refusal behavior was removed using several advanced techniques:
- Winsorization: Outlier activations were clipped at the 99th percentile to refine refusal direction estimation.
- Null-Space Projection: This method constrains weight updates to the null space of preservation activations, effectively maintaining the model's original capabilities.
- Adaptive Weighting: Gaussian-weighted per-layer ablation strength was applied, focusing on middle-to-later layers where refusal tendencies are more concentrated.
- Norm Preservation: The Frobenius norms of the weight matrices were preserved after projection, ensuring structural integrity.
Key Characteristics
- Uncensored Output: Designed to produce outputs without refusal behaviors, offering unrestricted responses.
- Vision-Language Capabilities: Inherits the multimodal capabilities of the base Qwen3-VL-8B-Instruct model.
- Technical Foundation: Leverages research from "Norm-Preserving Biprojected Abliteration" by Jim Lai and "AlphaEdit: Null-Space Constrained Knowledge Editing" by Fang et al., among others.
Use Cases
This model is intended for research and educational purposes where the removal of refusal behaviors is desired. Users should exercise responsibility, as the model will generate uncensored content.