jwest33/qwen3.5-9b-jspace-abliterated
jwest33/qwen3.5-9b-jspace-abliterated is a 9 billion parameter language model based on Qwen/Qwen3.5-9B, developed by jwest33. This model has undergone "J-space abliteration," a Jacobian-Lens-guided method that removes refusal behaviors by restricting the refusal direction to causally-active workspaces at each layer. It is designed to produce uncensored outputs, making it suitable for applications requiring unrestricted text generation, while maintaining the base model's hybrid attention architecture and 32768 token context length.
Loading preview...
Model Overview
jwest33/qwen3.5-9b-jspace-abliterated is a 9 billion parameter model derived from Qwen/Qwen3.5-9B, specifically modified to remove refusal behaviors. This modification, termed J-space abliteration, employs a sophisticated Jacobian-Lens-guided method to identify and restrict the refusal direction within the model's causally-active workspace at each layer. The process ensures that only components directly contributing to refusal tokens are ablated, preserving general helpfulness.
Key Techniques & Features
- J-Space Restricted Direction (Jacobian Lens): Utilizes
J_ℓ = E[∂h_final/∂h_ℓ]to build a low-rank subspace for refusal deliberation, projecting the mean-difference refusal direction onto this per-layer basis. - Data-Driven Concept Mining (Consensus): Empirically discovers refusal concepts from the model's own next-token distribution on harmful prompts, adapting to Qwen3.5's specific refusal vocabulary.
- Hybrid Architecture-Aware Intervention: Accounts for Qwen3.5-9B's interleaved full and linear attention layers, applying different ablation weights (1.0x for full, 0.4x for linear) and skipping recurrent dynamics projections.
- Projected Refusal Direction (GrimJim's Method): Orthogonalizes the raw refusal direction against the harmless mean to isolate the mechanistically specific refusal component.
- Norm Preservation: Maintains the original Frobenius norms of weight matrices post-projection.
Base Architecture
The underlying Qwen3.5-9B features a hybrid attention vision-language architecture with 32 layers, including 8 full attention layers and 24 linear attention layers. It has a hidden size of 4096, 16 attention heads, 4 KV heads, and supports a maximum context length of 262144 tokens.
Important Note
This model is designed to produce uncensored outputs. Users are advised to use it responsibly and ensure compliance with applicable laws and ethical standards.