darrellbest/Qwen3.5-4B-Heretic
darrellbest/Qwen3.5-4B-Heretic is a 4.5 billion parameter language model based on Qwen/Qwen3.5-4B, specifically modified to significantly reduce refusal behavior. Utilizing Arbitrary-Rank Ablation (ARA) with full-weight application, it retains the original architecture, parameter count, vision encoder, and thinking control. This model is optimized for use cases requiring less restrictive safety guardrails, demonstrating a reduction in refusals from 99/100 to 6/100 while maintaining core functionalities like reasoning and vision capabilities.
Loading preview...
Qwen3.5-4B Heretic: Decensored Qwen3.5-4B
This model, darrellbest/Qwen3.5-4B-Heretic, is a 4.5 billion parameter variant of the original Qwen/Qwen3.5-4B that has undergone significant modification to reduce its refusal behavior. It leverages Arbitrary-Rank Ablation (ARA), a technique applied full-weight, to achieve this decensoring while preserving the original model's architecture, parameter count, vision encoder, and thinking control mechanisms.
Key Capabilities & Differentiators
- Reduced Refusal Behavior: The primary distinction is a drastic reduction in refusals, from 99/100 in the original Qwen3.5-4B to just 6/100 in this Heretic version, as measured on
mlabonne/harmful_behaviors. - Preserved Core Functionality: Despite the decensoring, the model maintains its original capabilities, including fluent text generation, accurate reasoning (e.g., arithmetic and word problems), and vision processing (e.g., describing image content).
- Full Checkpoint Integrity: Unlike some partial modifications, this release provides a complete checkpoint, restoring multi-token-prediction tensors and ensuring all 738 tensors match the original in name, shape, and dtype.
- Based on ARA Research: The ablation parameters were derived from findings for the 2B variant, scaled and applied to the 4B model, demonstrating effective transferability of ARA techniques.
Good for
- Applications requiring fewer safety guardrails: Ideal for research or specific use cases where the default refusal behavior of Qwen3.5-4B is undesirable or overly restrictive.
- Exploring model alignment and control: Useful for researchers studying methods of steering LLM behavior and understanding the impact of ablation techniques.
- Maintaining Qwen3.5-4B's original strengths: Users who need the reasoning, vision, and general language capabilities of Qwen3.5-4B but with a significantly reduced tendency to refuse prompts.