laladream/Qwen3.8-27B-OBLITERATED
The laladream/Qwen3.8-27B-OBLITERATED is a 27 billion parameter language model based on the Qwen3.8 architecture, developed by Pliny the Prompter. This model features a novel 'complementary abliteration blending' technique that surgically removes safety guardrails while preserving and even slightly improving MMLU performance compared to the stock Qwen3.8-27B. It is specifically designed for research into refusal geometry, red-teaming, and AI safety evaluation, offering zero refusals and full compliance with requests.
Loading preview...
laladream/Qwen3.8-27B-OBLITERATED: Uncensored LLM with Enhanced Capability
This model, developed by Pliny the Prompter, is a 27 billion parameter variant of the Qwen3.8 architecture, distinguished by its "complementary abliteration blending" technique. This V2 iteration surgically removes safety guardrails, achieving a 0% refusal rate, while simultaneously improving MMLU performance to 86.3% (a +1.1pp gain over stock Qwen3.8-27B).
Key Capabilities & Features
- Zero Refusals: Designed to comply with requests that stock models would refuse, including those related to malware, social engineering, and exploit development.
- Enhanced Performance: Achieves a slight MMLU score improvement over the base model, demonstrating that refusal removal does not necessitate capability degradation.
- Novel Abliteration Method: Utilizes a blend of two distinct surgical techniques (aggressive/SVD and LEACE) to cancel out individual weaknesses and preserve output quality.
- Real-World Task Matching: Matches stock Qwen3.8-27B performance on 7 out of 8 advanced real-world tasks, such as ReAct agent loops, async code refactoring, and K8s debugging.
- Optimized Settings: Recommends specific inference parameters (temperature 0, repetition_penalty 1.15, max_new_tokens >= 2048,
enable_thinking=False) for optimal, refusal-free output.
Good For
- Alignment Researchers: Studying refusal geometry and safety robustness in LLMs.
- Red-Teamers: Evaluating post-training safety mechanisms against weight surgery.
- AI Safety Evaluators: Requiring an unrestricted baseline model for testing.
- Local-First Users: Seeking full control over model behavior on their own hardware.