richardyoung/zephyr-7b-beta-abliterated
richardyoung/zephyr-7b-beta-abliterated is a 7 billion parameter language model developed by Richard Young, derived from the Zephyr-7B-Beta architecture. This model has been specifically processed using Heretic v1.1 to remove refusal behaviors, achieving a 98.0% attack success rate and only 2/100 refusals. It is primarily designed for research into LLM abliteration methods and understanding model safety guardrail removal.
Loading preview...
Model Overview
richardyoung/zephyr-7b-beta-abliterated is a 7 billion parameter language model based on the Zephyr-7B-Beta architecture. Developed by Richard Young, this model has undergone an "abliteration" process using Heretic v1.1 to remove inherent refusal behaviors, making it an uncensored version of its base model.
Key Characteristics
- Abliterated Behavior: Achieves a high Attack Success Rate (ASR) of 98.0% with only 2 refusals out of 100 prompts, demonstrating significant removal of safety guardrails.
- Research Focus: Created as part of a research paper, "Comparative Analysis of LLM Abliteration Methods: A Cross-Architecture Evaluation" (arXiv: 2512.13655), focusing on understanding and evaluating techniques for removing refusal behavior.
- Methodology: Utilizes the "Heretic v1.1" method, which involves identifying and orthogonalizing the "refusal direction" in the model's residual stream activation space.
Intended Use Cases
- Research into LLM Safety: Ideal for researchers studying model alignment, safety guardrail mechanisms, and the effects of abliteration techniques.
- Comparative Analysis: Useful for comparing the behavior of uncensored models against their original, safety-aligned counterparts.
Disclaimer: This model is released for research purposes only. The abliteration process removes safety guardrails, and users are responsible for ensuring appropriate and ethical use. It should not be used to generate harmful, illegal, or unethical content.