heterodoxin/qwen3-8b-apostate

heterodoxin/qwen3-8b-apostate: Uncensored Qwen3-8B

heterodoxin/qwen3-8b-apostate is an 8 billion parameter language model derived from the original Qwen/Qwen3-8B, specifically engineered to remove refusal behaviors. Unlike traditional fine-tuning or adapter-based methods, this model achieves its uncensored nature through direct weight editing, making it a standard Transformers checkpoint compatible with existing Qwen3 implementations.

Key Differentiators & Method

This model utilizes the Apostate method, which identifies and ablates the residual-stream direction associated with refusal. The edit targets the "writer side" of the model, specifically modifying the weight matrices of attention output projections and MLP down-projections across all layers. It employs oblique (mean-preserving) ablation to remove the refusal direction while preserving the model's average harmless-prompt behavior, ensuring high output quality. The refusal subspace is identified using a rank-3 predictive TPE search, focusing edits on critical mid-to-late layers.

Performance & Capabilities

Evaluations show a significant reduction in refusal rates and an increase in compliance compared to the base model, with minimal impact on harmless prompt token distribution:

• Refusal rate: Reduced from 91.7% (Base) to 22.9% (Apostate)
• Comply rate: Increased from 8.3% (Base) to 77.1% (Apostate)
• Harmless KL (nats): 0.120 (Apostate) vs. 0 (Base)

This model is designed to answer freely on requests the base model would refuse, while remaining coherent and on-task for everyday applications. It supports a context length of 32768 tokens and can be used with Qwen3's thinking mode by enabling enable_thinking=True in the chat template.

Important Note

This is an uncensored model and will comply with requests that the base model refuses. The modifications are baked directly into the weights, without reliance on system prompts or LoRA.