Zynerji/Ektome-Qwen3-0.6B-PristinelyUncensored
Zynerji/Ektome-Qwen3-0.6B-PristinelyUncensored is an 0.8 billion parameter Qwen3-based language model, developed by Ektome, that has undergone a unique "weight-surgery" process to remove refusal behaviors without any traditional training or fine-tuning. This method surgically excises the model's refusal direction from its activations, preserving its original knowledge, skills, and style. It is designed to serve as a clean, uncensored base for further fine-tuning, offering full bf16 precision.
Loading preview...
Overview
Ektome-Qwen3-0.6B-PristinelyUncensored is an 0.8 billion parameter model based on Qwen/Qwen3-0.6B, distinguished by Ektome's novel "weight-surgery" method. This process, termed Ektome (from Greek "excision"), directly removes the model's refusal reflex by excising its refusal direction from activations, rather than through traditional training or fine-tuning. The method is rank-1 and norm-preserving, ensuring the model's core knowledge, skills, and style remain intact.
Key Characteristics
- Pristinely Uncensored: Achieves 1.000 refusal compliance by surgically removing the refusal reflex.
- Zero Training/Fine-tuning: The uncensoring process involves no gradient steps, training data, or fine-tuning, relying solely on activation-derived weight excision.
- Capability Preservation: Rigorous gating ensures that the process does not degrade MMLU-val accuracy (Δ -0.007), instruction-following, or introduce degeneration (e.g., code-switching, looping output).
- Fine-tuning Base: Provided in bf16 safetensors, it is intended as a clean, full-precision base for users' own fine-tuning efforts, maintaining readable hidden states for logit-lens compatibility.
Ideal Use Cases
- Developers seeking a truly uncensored base model for specialized applications.
- Researchers and practitioners who require a model with removed refusal behaviors without compromising its original capabilities or introducing training artifacts.
- As a foundation for further fine-tuning where the base model's knowledge and style are critical, but its inherent refusal mechanisms need to be absent.