richardyoung/DeepSeek-R1-Distill-Qwen-7B-abliterated-obliteratus
The richardyoung/DeepSeek-R1-Distill-Qwen-7B-abliterated-obliteratus is a 7.6 billion parameter language model, based on the DeepSeek-R1-Distill-Qwen-7B architecture, developed by Richard Young. This model has undergone an "abliteration" process using the OBLITERATUS method to significantly reduce refusal behavior, achieving a 50% attack success rate. It is specifically designed for research into uncensored model behavior and the study of refusal direction orthogonalization techniques in LLMs. The model maintains a 32768 token context length.
Loading preview...
Model Overview
This model, richardyoung/DeepSeek-R1-Distill-Qwen-7B-abliterated-obliteratus, is a 7.6 billion parameter variant of the DeepSeek-R1-Distill-Qwen-7B architecture. Developed by Richard Young, its primary distinction is the application of an advanced "abliteration" technique called OBLITERATUS. This process aims to remove refusal behavior by identifying and orthogonalizing the "refusal direction" within the model's residual stream activation space.
Key Characteristics
- Abliterated (Uncensored): Specifically modified to reduce refusal behavior, achieving a 50% Attack Success Rate (ASR) and 50/100 refusals in testing.
- Research-Focused: Created as part of a research paper, "Comparative Analysis of LLM Abliteration Methods: Scaling to MoE Architectures and Modern Tools" (arXiv: 2512.13655).
- Methodology: Utilizes the OBLITERATUS (advanced) method for abliteration, a technique for removing refusal behavior from language models.
- Context Length: Supports a context length of 32768 tokens.
Intended Use Cases
- Research into LLM Safety & Alignment: Ideal for academic and research purposes to study model refusal mechanisms and abliteration techniques.
- Understanding Model Biases: Useful for investigating how models respond without inherent safety guardrails.
- Comparative Analysis: Can be used to compare the effectiveness of different abliteration methods.
Disclaimer: This model is released for research purposes only. The abliteration process removes safety guardrails, and users are responsible for ensuring appropriate use. It should not be used to generate harmful, illegal, or unethical content.