Zynerji/Ektome-Qwen2-1.5Bi-PristinelyUncensored
Zynerji's Ektome-Qwen2-1.5Bi-PristinelyUncensored is a 1.5 billion parameter Qwen2-based causal language model with a 32768 token context length. This model has undergone a unique 'weight-surgery' process called Ektome, which removes refusal tendencies directly from its activations without any training or fine-tuning. It is designed to be a clean, uncensored base model for further fine-tuning, maintaining its original knowledge and skills while eliminating refusal behaviors.
Loading preview...
Ektome-Qwen2-1.5Bi-PristinelyUncensored Overview
This model, developed by Zynerji, is a 1.5 billion parameter variant of the Qwen2-1.5B-Instruct architecture. Its core innovation lies in the Ektome method, a "weight-surgery" technique that directly excises the model's refusal direction from its activations. This process involves zero training data, zero gradient steps, and zero fine-tuning, ensuring that the model's original knowledge, skills, and stylistic capabilities remain intact.
Key Characteristics & Differentiation
- Pristinely Uncensored: Achieves a refusal compliance rate of 0.990, effectively removing refusal behaviors without traditional fine-tuning.
- Weight-Surgery Method: Utilizes a novel approach that modifies the model's weights by reading and excising refusal directions, rather than through data-driven training.
- Capability Preservation: Rigorous gating ensures that the removal of refusal does not degrade MMLU accuracy (Δ -0.005), instruction-following, or introduce generation issues like code-switching or degeneration.
- Clean Fine-tuning Base: Provided as bf16 safetensors, these weights are intended as a robust and uncensored foundation for subsequent fine-tuning by users.
Ideal Use Cases
This model is particularly well-suited for developers and researchers who require:
- A highly compliant and uncensored base model for various applications.
- A foundation for custom fine-tuning where the original model's knowledge and skills are critical, but refusal behaviors need to be absent.
- Exploration of novel model modification techniques that bypass traditional training paradigms.