Zynerji/Ektome-Phi-3-mini-4k-instruct-PristinelyUncensored
Zynerji/Ektome-Phi-3-mini-4k-instruct-PristinelyUncensored is a 4 billion parameter instruction-tuned language model based on the Phi-3-mini architecture, featuring a 4096-token context length. Developed by Zynerji, this model is specifically engineered using the Ektomē method to be uncensored by design, isolating and removing only refusal-specific components without impacting general helpfulness or knowledge. It is intended for use cases requiring an uncensored model where the user is accountable for its outputs.
Loading preview...
Ektome-Phi-3-mini-4k-instruct-PristinelyUncensored Overview
This model, developed by Zynerji, is an uncensored version of the Phi-3-mini-4k-instruct, utilizing a proprietary method called Ektomē (excision). The Ektomē process aims to remove only refusal-specific components from the model, preserving its general helpfulness and reasoning capabilities without additional training or distillation. This approach contrasts with standard "abliteration" methods that can inadvertently remove knowledge and reasoning alongside refusal directions.
Key Characteristics
- Uncensored by Design: Engineered to not refuse, providing direct responses to user prompts.
- Ektomē Method: Employs a proprietary excision operator and depth-selection procedure to isolate and remove refusal-specific components norm-preservingly.
- Pristine Model Basis: The excision is applied directly to the pristine model, avoiding damage that would need repair.
- No Capability-Retention Claim: While designed to retain capabilities, no n=2800 certificate has been run, meaning no formal capability-retention claim is made, and any reported numbers are point estimates without confidence intervals.
- 4 Billion Parameters: A compact model size for efficient deployment.
- 4096-token Context Length: Supports processing of moderately long inputs.
Intended Use
This model is suitable for applications where an uncensored language model is explicitly desired, and the user assumes full accountability for the generated content. It is important to note that the model's certificate bounds only capability retention (though not formally certified for this version) and does not certify safety, factual accuracy, or fitness for any specific purpose. Users should be aware that it will not refuse, even for potentially harmful or inappropriate queries.