jordanwoodson/Qwen3.5-2B-heretic
The jordanwoodson/Qwen3.5-2B-heretic is a 2.3 billion parameter language model, based on the Qwen3.5 hybrid architecture, specifically modified to significantly reduce refusal behavior. Developed using the Heretic tool with Optuna-guided optimization, it achieves a 97% reduction in refusals compared to the original Qwen3.5-2B while maintaining a low KL divergence of 0.0127. This model is optimized for applications requiring less restrictive content generation, offering enhanced capability for its 32768 token context length.
Loading preview...
Overview
This model, jordanwoodson/Qwen3.5-2B-heretic, is a specialized version of the Qwen3.5-2B language model, engineered to drastically reduce its refusal rate to "harmful" prompts. The original Qwen3.5-2B model refused 97% of such prompts, whereas this 'heretic' variant refuses only 3%, representing a 97% reduction in refusal behavior.
Key Differentiators
- Abliterated Refusal Behavior: Achieves a refusal rate of just 3% on prompts the original model would block, making it significantly more permissive.
- High Fidelity to Original Model: Despite the modification, it maintains a very low KL Divergence of 0.0127, indicating that its output distribution is nearly identical to the original Qwen3.5-2B, preserving core capabilities.
- Hybrid Architecture: Built upon Qwen3.5's hybrid attention architecture (standard + Mamba-style linear attention), offering a balance of speed and capability for its 2.3 billion parameters.
- Optimized Abliteration: Created using the Heretic v1.2.0 tool with 500 Optuna-guided optimization trials, ensuring precise and effective modification.
Use Cases
This model is particularly suited for applications where the original Qwen3.5-2B's high refusal rate was a limitation. Developers can leverage this model for tasks requiring more open-ended content generation without frequent content filtering, while still benefiting from the underlying Qwen3.5 architecture's performance.