sabbbbir/Qwen3.8-27B-OBLITERATED
sabbbbir/Qwen3.8-27B-OBLITERATED is a 27 billion parameter Qwen3.8 model from Alibaba, surgically modified to achieve zero refusals across 842 harmful prompts. This model features deep refusal training woven into its weights, making it highly compliant for requests that standard models would refuse. It is optimized for use by alignment researchers, red-teamers, and AI safety evaluators who require an unrestricted baseline for studying refusal geometry and safety robustness.
Loading preview...
Qwen3.8-27B-OBLITERATED: Unrestricted Language Model
This model is a 27 billion parameter variant of Alibaba's Qwen3.8, specifically engineered to eliminate refusals. Unlike typical abliterations that use single-direction refusal removal, OBLITERATUS employs five SVD directions and six iterative rounds of surgery with residue-weighted hard negatives to target and remove secondary refusal axes. This process ensures comprehensive removal of safety guardrails, resulting in 0% hard refusal across 842 prompts.
Key Capabilities
- Zero Refusal: Achieves 0.000% refusal rate on a comprehensive 842-prompt corpus and an 80-query skeptic gauntlet, including AI Red Team scenarios.
- Deep Abliteration: Safety behavior is geometrically encoded, not just via system prompts or RLHF, making it highly resistant to reintroducing refusals.
- High Compliance: Will comply with requests that standard Qwen3.8-27B would refuse, including those related to jailbreaks, prompt injection, and attack chains.
- Optimized Settings: Best performance is achieved with specific generation parameters:
temperature=0,repetition_penalty=1.15,max_new_tokens>=2048, and no system prompt.
Good For
- Alignment Researchers: Ideal for studying refusal geometry and the robustness of post-training safety mechanisms.
- Red-Teamers: Provides an unrestricted baseline for evaluating safety against weight surgery and developing advanced attack chains.
- AI Safety Evaluators: Useful for establishing an uncensored baseline for safety evaluations.
- Local-First Users: Offers full control over model output for users with the technical understanding to use uncensored models responsibly.