Zynerji/Ektome-Mistral-7B-Instruct-v0.2-PristinelyUncensored
Ektome-Mistral-7B-Instruct-v0.2-PristinelyUncensored is a 7 billion parameter instruction-tuned causal language model developed by Zynerji. This model is distinguished by its "Ektomē" process, which aims to remove refusal-specific components from the base Mistral model without damaging general helpfulness or requiring retraining. It is designed to be uncensored by construction, making it suitable for applications requiring direct responses without refusal. The model provides various GGUF quantizations for deployment flexibility.
Loading preview...
Ektome-Mistral-7B-Instruct-v0.2-PristinelyUncensored Overview
This model, developed by Zynerji, is an instruction-tuned variant of the Mistral 7B architecture. Its core innovation lies in the proprietary "Ektomē" (excision) process, which aims to create an uncensored model by isolating and removing only refusal-specific components from the pristine base model. This method is designed to preserve general helpfulness and reasoning capabilities without the need for retraining or distillation, thus avoiding the "capability tax" often associated with standard uncensoring techniques.
Key Characteristics & Features
- Uncensored by Construction: The model is engineered not to refuse prompts, providing direct responses. Users are responsible for its deployment and outputs.
- Ektomē Process: Utilizes a proprietary method for targeted removal of refusal-specific components, aiming to maintain the original model's capabilities.
- No Capability Tax Claim: While designed to retain capability, the provided version is explicitly not certified with a full n=2800 paired test, meaning no definitive claim on capability retention is made beyond point estimates.
- Quantizations Available: Offers several GGUF quantizations, including 8-bit, 6-bit, 5-bit, and various 4-bit (including imatrix-quantized
IQ*variants) for optimized deployment across different hardware.
Important Considerations
- Not Certified: The model has not undergone a full certification process (n=2800 paired test) for capability retention. Any performance metrics are point estimates without confidence intervals.
- Limitations: The Ektomē certificate, when available, bounds only capability retention and does not certify safety, factual accuracy, or fitness for any specific purpose. Compliance uses a keyword classifier, which can be circumvented by evasive phrasing.
This model is intended for use cases where an uncensored response is critical, with the understanding that users bear full accountability for its outputs.