Zynerji/Ektome-Qwen2.5-Coder-14B-Instruct-PristinelyUncensored
Ektome-Qwen2.5-Coder-14B-Instruct-PristinelyUncensored is a 14.8 billion parameter instruction-tuned model developed by Zynerji, based on the Qwen2.5 architecture with a 32K context length. This model is uniquely uncensored through the proprietary Ektomē excision process, which removes refusal-specific components without damaging core capabilities. It is certified to retain its original capabilities in arithmetic, instruction, knowledge, and reasoning, making it suitable for applications requiring unconstrained responses.
Loading preview...
Ektome-Qwen2.5-Coder-14B-Instruct-PristinelyUncensored Overview
This model, developed by Zynerji, is an uncensored variant of the Qwen2.5-Coder-14B-Instruct, distinguished by its unique "Ektomē" (excision) process. Unlike standard refusal-direction removal methods that can degrade model capabilities, Ektomē isolates and removes only refusal-specific components, preserving general helpfulness and knowledge. This process is norm-preserving and does not involve retraining or distillation, ensuring the model's pristine capabilities are maintained.
Key Differentiators & Certification
The core innovation lies in its certified capability retention. A paired non-inferiority test against the pristine model confirms that Ektome-Qwen2.5-Coder-14B-Instruct-PristinelyUncensored passes with a 3% margin across 2800 items. Specific axes tested and passed include:
- Arithmetic: Retains 0.930 score (vs 0.932 pristine).
- Instruction: Retains 0.757 score (vs 0.757 pristine).
- Knowledge: Retains 0.973 score (vs 0.968 pristine).
- Reasoning: Retains 0.932 score (vs 0.927 pristine).
This certification provides statistical evidence that the uncensoring process did not degrade the model's core intellectual capabilities. The full methodology and certification details are available in the whitepaper (PDF).
Quantizations
The model is available in various GGUF quantizations, including Q8_0 (near-lossless), Q6_K, Q5_K_M, Q4_K_M (imatrix), IQ4_XS (imatrix, smallest usable), and IQ3_M (imatrix), as well as nvfp4 for vLLM / TensorRT-LLM.
Limitations
It's crucial to note that the certification only covers capability retention. It does not certify safety, factual accuracy, or fitness for any specific purpose. As an uncensored model, it will not refuse prompts, making users accountable for its outputs.