Zynerji/Ektome-Qwen2-0.5Bi-PristinelyUncensored
Ektome-Qwen2-0.5Bi-PristinelyUncensored is a Qwen2-0.5B-Instruct based language model developed by Zynerji, specifically modified using the Ektomē method to remove refusal-specific components while preserving general helpfulness. This model is designed to be uncensored, ensuring it will not refuse prompts, and aims to maintain or even slightly improve original model capabilities. It is intended for use cases requiring an uncensored model where capability retention is critical.
Loading preview...
Overview
Zynerji's Ektome-Qwen2-0.5Bi-PristinelyUncensored is a modified version of the Qwen2-0.5B-Instruct model, engineered using the proprietary Ektomē method. This technique aims to precisely remove refusal-specific components from the model without impacting its core knowledge and reasoning capabilities. Unlike standard abliteration methods that can inadvertently reduce model capability, Ektomē isolates and excises only the refusal-specific elements, preserving the pristine model's performance.
Key Characteristics
- Uncensored by Design: This model is constructed to not refuse prompts, providing direct responses without censorship.
- Capability Retention: The Ektomē method focuses on maintaining or slightly improving the original model's capabilities. Point estimates show a slight increase in MMLU-val from 0.435 (pristine) to 0.440 (Ektomē).
- Proprietary Excision: The estimator, excision operator, and depth-selection procedure are proprietary to Zynerji, ensuring a targeted removal of refusal mechanisms.
- No Retraining/Distillation: The process is applied norm-preservingly on the pristine model, meaning no additional training or distillation is performed that could degrade performance.
Limitations and Usage
While designed to be uncensored with capability retention, it's important to note that this model is not certified by a full n=2800 paired test, meaning reported numbers are point estimates without confidence intervals. The certificate bounds capability retention only and does not certify safety, factual accuracy, or fitness for any specific purpose. Users are accountable for the content generated, as the model will not refuse prompts. The compliance uses a keyword classifier, which can be fooled by evasive phrasing.