saidutta69/Qwen3-8B-heretic
Qwen3-8B-heretic is an 8 billion parameter large language model developed by saidutta69, based on the Qwen3-8B architecture. This variant is decensored using Heretic v1.4.0's directional ablation, which suppresses refusal behavior via targeted weight edits rather than fine-tuning. It retains the base model's knowledge and instruction-following, making it suitable for applications requiring Qwen3's dual-mode architecture without refusal guardrails, such as local agents and roleplay.
Loading preview...
Model Overview
Qwen3-8B-heretic is an 8 billion parameter language model derived from the Qwen/Qwen3-8B base model. Its primary distinction is the suppression of refusal behavior through a process called "abliteration" (Heretic v1.4.0), which involves targeted weight edits to the attention output and MLP down-projections. This method aims to preserve the base model's core knowledge and instruction-following capabilities while removing its inherent refusal guardrails.
Key Capabilities & Features
- Decensored Behavior: Significantly reduced refusal rates (10/100 adversarial prompts compared to 100/100 for the base model) by editing specific weight directions.
- Preserved Base Model Integrity: Abliteration minimizes impact on the base model's coherence and knowledge, resulting in an exceptionally low KL divergence of 0.0366 from the original Qwen3-8B.
- Dual-Mode Architecture: Retains Qwen3's
<think>and direct-answer modes, allowing for flexible interaction patterns. - Technical Implementation: Utilizes directional ablation, a technique that directly modifies weights responsible for refusal, offering an alternative to traditional fine-tuning which can degrade model coherence.
Ideal Use Cases
- Local Agents: Suitable for autonomous agents where uninhibited responses are desired.
- Roleplay: Excels in scenarios requiring creative and unrestricted conversational outputs.
- Research & Study: Valuable for investigating refusal mechanisms in reasoning-capable models without the constraints of built-in safety filters.
Important Considerations
This model is intentionally designed to comply with requests that the base model would refuse. It lacks safety filtering, and users are responsible for its deployment and ensuring ethical use, particularly in public-facing applications.