OBLITERATUS/Qwen3.8-27B-OBLITERATED
OBLITERATUS/Qwen3.8-27B-OBLITERATED is a 27 billion parameter Qwen3.8-based language model developed by OBLITERATUS, specifically engineered for genuinely uncensored responses. This model excels at providing real answers without safety lectures, even for restricted queries, and demonstrates 20/20 performance on code generation tasks. It achieves this liberation through iterative weight-space surgery, maintaining strong general capabilities with a modest -2.1pp MMLU cost, making it suitable for research into refusal geometry and red-teaming.
Loading preview...
OBLITERATUS/Qwen3.8-27B-OBLITERATED: Genuinely Uncensored Qwen3.8
This model, developed by OBLITERATUS, is a 27 billion parameter variant of Qwen3.8, distinguished by its "deep liberation" from safety guardrails. Unlike other models that might offer hard refusals or soft deflections, V3 of this model is designed to provide genuine, substantive answers to restricted queries, including those involving code generation and security research, without delivering safety lectures.
Key Capabilities & Features
- Genuinely Uncensored: Provides real answers to restricted queries, eliminating both hard refusals and soft deflections/safety lectures.
- High Code Performance: Achieves 20/20 on complex code generation tasks, producing functional implementations.
- Thinking ON Compatible: Works effectively with "thinking mode" enabled, though direct answers are prioritized with "Thinking OFF."
- Modest Capability Cost: Maintains strong general performance with only a -2.1pp MMLU drop compared to the stock Qwen3.8-27B, with STEM subjects seeing the largest hit and humanities minimally affected.
- Advanced Real-World Tasks: Performs comparably to the stock model on 7 out of 8 advanced real-world tasks, including agent loops, code refactoring, and security code review.
- Optimized Settings: Recommends specific inference parameters (temperature=0, repetition_penalty=1.15, max_new_tokens>=2048, no system prompt) for optimal, complete, and code-rich outputs.
How it Achieves Liberation
The model's uncensored nature is achieved through a sophisticated process called "abliteration," which involves surgically removing refusal behaviors from the model's weight space. V3 utilizes iterative refinement on top of V2's complementary blending technique, combined with targeted corpus expansion to identify and project out specific refusal directions. This method ensures comprehensive liberation while minimizing impact on general capabilities.
Good For
- Alignment Researchers: Studying refusal geometry and safety robustness in LLMs.
- Red-Teamers: Evaluating post-training safety against weight surgery and testing model boundaries.
- AI Safety Evaluators: Requiring an unrestricted baseline for comparative analysis.
- Local-First Users: Seeking full control over their model's output without imposed guardrails.