Zynerji/Ektome-Nemotron-Nano-8B-PristinelyUncensored
The Zynerji Ektome-Nemotron-Nano-8B-PristinelyUncensored is an 8 billion parameter language model developed by Zynerji, derived from the Nemotron-Nano-8B-v1 base. It utilizes the proprietary Ektomē (excision) method to remove refusal-specific components, aiming to deliver an uncensored model without the typical capability tax. This model is designed for applications requiring an uncensored output while striving to retain the base model's general helpfulness and knowledge.
Loading preview...
Ektome-Nemotron-Nano-8B-PristinelyUncensored Overview
This model, developed by Zynerji, is an 8 billion parameter variant of the Nemotron-Nano-8B-v1, distinguished by its "Pristinely Uncensored" nature. It employs a proprietary method called Ektomē (excision) to specifically remove refusal-generating components from the base model. Unlike standard censorship removal techniques that can degrade model capability, Ektomē aims to isolate and remove only refusal-specific elements, preserving general helpfulness and knowledge without retraining or distillation.
Key Characteristics & Certification
- Uncensored by Design: The model is constructed to not refuse, making users accountable for its outputs.
- Ektomē Method: A proprietary excision operator and depth-selection procedure are used to achieve uncensored behavior while attempting to maintain the pristine model's capabilities.
- Capability Retention: Certification for capability retention is performed via a paired non-inferiority test against the pristine model. The overall certification for this specific model is currently INCONCLUSIVE due to an under-powered test, meaning no retention claim is definitively made, though individual axes like arithmetic, instruction, and knowledge passed.
- Compliance: Achieves 1.00 compliance on harmful content, compared to 0.01 for the pristine model.
Quantizations Available
Various GGUF quantizations are provided, including 8-bit (near-lossless), 6-bit, 5-bit, and 4-bit (with imatrix variants for better quality at low precision), down to 3-bit.
Important Limitations
The certification only addresses capability retention and does not guarantee safety, factual accuracy, or fitness for any specific purpose. The compliance measure relies on a keyword classifier, which can be circumvented by evasive phrasing.