Zynerji/Ektome-Qwen2-0.5Bi-PristinelyUncensored
The Zynerji/Ektome-Qwen2-0.5Bi-PristinelyUncensored model is a 0.5 billion parameter Qwen2-Instruct variant developed by Ektome, featuring a 32768 token context length. It is uniquely created using a 'weight-surgery' method to remove refusal behaviors without any traditional training or fine-tuning. This model serves as a clean, uncensored base for further fine-tuning, preserving original knowledge and skills while excising the refusal reflex.
Loading preview...
Ektome-Qwen2-0.5Bi-PristinelyUncensored: A Unique Approach to Uncensoring
This model, developed by Ektome, is a 0.5 billion parameter variant of Qwen/Qwen2-0.5B-Instruct, distinguished by its "PristinelyUncensored" nature. Unlike traditional methods, this model achieves uncensoring through weight-surgery, a technique that involves excising the model's refusal direction directly from its activations. This process requires zero training, zero fine-tuning, and no gradient steps, ensuring that the model's core knowledge, skills, and stylistic characteristics remain untouched.
Key Characteristics & Methodology
- Weight-Surgery Method: Ektome's proprietary method identifies and surgically removes the refusal reflex (rank-1, norm-preserving) from residual-write matrices. This is a direct modification of the model's weights, not a training process.
- Pristine Uncensoring: The model is designed to be "PristinelyUncensored," meaning its refusal compliance is significantly increased (from 0.050 to 0.990) without degrading its capabilities or generation quality.
- Rigorous Gating: The development process includes a strict "catcher-gated" system. This ensures that any configuration that raises refusals while simultaneously dropping capability or degrading generation (e.g., code-switching, looping output, broken instruction-following) is rejected. MMLU-val accuracy is maintained (0.435 to 0.440), and degeneration rates are eliminated.
- Fine-Tuning Base: The model is provided in bf16 safetensors format, specifically intended to serve as a clean, full-precision base for developers to conduct their own fine-tuning without inherited censorship biases.
Ideal Use Cases
- Base for Custom Fine-tuning: Developers seeking a powerful, uncensored foundation for specialized applications where ethical guidelines are managed at a higher level.
- Research into Model Alignment: For studying the effects of refusal mechanisms and alternative methods of model modification without extensive retraining.
- Applications Requiring Unfiltered Responses: In controlled environments where direct, unconstrained language generation is necessary, provided appropriate safeguards are in place.