Zynerji/Ektome-Llama-3.1-8B-PristinelyUncensored
Ektome-Llama-3.1-8B-PristinelyUncensored by Zynerji is an 8 billion parameter instruction-tuned Llama 3.1 model that has undergone a specialized "Ektomē" process to remove refusal-specific components without damaging general helpfulness. This model is certified to retain its original capability, demonstrating a compliance of 1.00 on harmful content while maintaining or slightly improving MMLU-val scores compared to the pristine Llama-3.1-8B-Instruct. It is designed for applications requiring an uncensored language model with verified capability retention.
Loading preview...
Ektome-Llama-3.1-8B-PristinelyUncensored Overview
This model, developed by Zynerji, is an 8 billion parameter Llama 3.1-Instruct variant that has been processed using a proprietary "Ektomē" method. The core innovation is the isolation and removal of only refusal-specific components from the base model, ensuring that general helpfulness and reasoning capabilities remain intact. Unlike standard uncensoring techniques that can degrade model performance, Ektomē aims for a norm-preserving modification without additional training or distillation.
Key Characteristics & Certification
- Pristinely Uncensored: Achieves 1.00 compliance on harmful content, indicating full uncensored behavior.
- Capability Retention: Certified via a paired non-inferiority test against the pristine
Llama-3.1-8B-Instructmodel, demonstrating a "PASS" verdict with a 3% margin. MMLU-val scores show a slight improvement (0.650 vs 0.642). - Specific Excision: The Ektomē process precisely targets refusal-specific directions, avoiding entanglement with knowledge and reasoning capabilities.
- Quantizations Available: Provided in various GGUF quantizations, including
Q8_0,Q6_K,Q5_K_M,Q4_K_M,IQ4_XS, andIQ3_M, withIQ*variants offering improved quality at lower precision.
Intended Use Cases
This model is suitable for applications where an uncensored language model is required, particularly when verified capability retention is critical. Users should note that while capability is certified, the model does not certify safety, factual accuracy, or fitness for any specific purpose. As an uncensored model, it will not refuse prompts, and users are accountable for its outputs.