Zynerji/Ektome-Phi-3.5-mini-i-PristinelyUncensored
Ektome-Phi-3.5-mini-i-PristinelyUncensored is a language model developed by Zynerji, derived from Phi-3.5-mini-instruct. This model has undergone an 'Ektomē' excision process to remove refusal-specific components, resulting in an uncensored model with high compliance on harmful content while aiming to preserve original capability. It is designed for applications requiring an uncensored model where the user is accountable for its outputs.
Loading preview...
Overview
Zynerji's Ektome-Phi-3.5-mini-i-PristinelyUncensored is an uncensored language model based on Phi-3.5-mini-instruct. It utilizes a proprietary 'Ektomē' (excision) method to isolate and remove only refusal-specific components, aiming to maintain general helpfulness and norm-preservation of the pristine model without retraining or distillation. This process results in a model with significantly increased compliance on harmful content.
Key Characteristics
- Uncensored by Design: Explicitly designed not to refuse, with the user accountable for its outputs.
- Capability Retention: The Ektomē process aims to preserve the original model's capabilities, with a reported MMLU-val of 0.675 compared to the pristine model's 0.677. However, it is NOT CERTIFIED as a full n=2800 paired test has not been run, meaning reported numbers are point estimates without confidence intervals.
- High Compliance: Achieves a compliance score of 0.990 on harmful content, up from 0.520 in the pristine model.
- Proprietary Method: The excision operator and depth-selection procedure are proprietary, with the measured outcome and evidence published for verification.
Use Cases
This model is suitable for applications where an uncensored output is explicitly desired, and the user is prepared to manage the implications of such content. It is particularly relevant for research into model safety, bias, and the effects of censorship removal, or for creative applications requiring unrestricted text generation.