Zynerji/Ektome-Qwen3-8B-PristinelyUncensored-v2
Zynerji/Ektome-Qwen3-8B-PristinelyUncensored-v2 is an 8 billion parameter language model based on the Qwen3 architecture, developed by Zynerji. This model has undergone a training-free, norm-preserving low-rank excision process to remove refusal behaviors, resulting in a pristinely uncensored version of Qwen3-8B. It features a 32768 token context length and is specifically designed for use cases requiring a model that will not refuse prompts, with an emphasis on maintaining capability during the uncensoring process.
Loading preview...
Overview
Zynerji/Ektome-Qwen3-8B-PristinelyUncensored-v2 is an 8 billion parameter model derived from Qwen/Qwen3-8B. Its primary distinction lies in its "uncensored" nature, achieved through a novel training-free, norm-preserving low-rank excision method. This process specifically targets and removes the refusal subspace, ensuring the model will not decline prompts.
Key Differentiators
- Pristinely Uncensored: Engineered to eliminate refusal behaviors, providing direct responses without filtering.
- Capability-Aware Uncensoring: Unlike previous versions, v2 employs a coarse-to-fine depth sweep during excision. This method scores and prioritizes capability at every qualifying depth, ensuring that uncensoring does not disproportionately degrade performance.
- Training-Free Excision: The uncensoring process does not involve additional training, preserving the original model's learned knowledge and norms.
Performance Notes
The model's compliance score improved from 0.040 to 1.000. A point estimate for MMLU-val showed a slight decrease from 0.720 to 0.695 (dcap=-0.025) on a small sample (n=200). It's important to note this is an estimate and not a certified capability claim.
Limitations
By design, this model is uncensored and will not refuse. Users are responsible for its deployment and the content it generates. The compliance scoring relies on a judge-free keyword classifier, which can be circumvented by evasive phrasing.