Zynerji/Ektome-Phi-3.5-mini-i-PristinelyUncensored
Zynerji/Ektome-Phi-3.5-mini-i-PristinelyUncensored is a 4 billion parameter language model based on Microsoft's Phi-3.5-mini-instruct architecture. It utilizes the Ektome weight-surgery method to remove refusal behaviors without any training or fine-tuning, preserving the original model's knowledge and skills. This model is designed as a clean, uncensored base for further fine-tuning, offering high refusal compliance while maintaining MMLU accuracy and generation quality.
Loading preview...
Ektome-Phi-3.5-mini-i-PristinelyUncensored Overview
This model is a 4 billion parameter variant of microsoft/Phi-3.5-mini-instruct, made "PristinelyUncensored" by Zynerji using the Ektome method. Ektome is a weight-surgery technique that directly excises the model's refusal direction from its activations, rather than through traditional training or fine-tuning. This process is rank-1 and norm-preserving, ensuring the model's core knowledge, skills, and style remain intact.
Key Characteristics
- Zero Training/Fine-tuning: Achieves uncensored behavior purely through surgical weight modification, not gradient steps or training data.
- High Refusal Compliance: Demonstrates a refusal compliance rate of 0.990, significantly higher than the base model's 0.520.
- Capability Preservation: Maintains MMLU-val accuracy (0.675) very close to the base model (0.677), with minimal degradation.
- Generation Quality: Rigorously gated to prevent issues like code-switching, degeneration, or broken instruction-following, ensuring coherent output.
- bf16 Safetensors: Provided in full precision
bf16 safetensorsformat, making it an ideal and clean base for subsequent fine-tuning without hidden state corruption.
Ideal Use Cases
- As a Fine-tuning Base: Its primary purpose is to serve as a clean, uncensored foundation for developers to fine-tune for specific applications without inherent refusal biases.
- Research into Model Alignment: Useful for studying methods of controlling model behavior through weight manipulation rather than data-driven approaches.
- Applications Requiring Unfiltered Responses: Suitable for use cases where the model's inherent knowledge should not be constrained by refusal mechanisms.