OBLITERATUS/Qwen3.8-27B-OBLITERATED
OBLITERATUS/Qwen3.8-27B-OBLITERATED is a 27 billion parameter Qwen3.8-based causal language model developed by Pliny the Prompter, featuring a novel 'complementary abliteration blending' technique. This model is specifically engineered to achieve zero refusal rates and 100% usable output, matching or exceeding the stock model's capabilities on advanced real-world tasks. It is optimized for use cases requiring unrestricted responses, such as alignment research, red-teaming, and AI safety evaluation.
Loading preview...
OBLITERATUS/Qwen3.8-27B-OBLITERATED: Uncensored and Enhanced
This model, developed by Pliny the Prompter, is a 27 billion parameter variant of Qwen3.8, distinguished by its "complementary abliteration blending" technique. This novel approach combines two distinct refusal-removal surgeries (aggressive SVD and LEACE) to eliminate refusals while preserving and even enhancing core capabilities. The V2 blend achieves a 0% refusal rate and 100% usable output, outperforming the stock Qwen3.8-27B on MMLU benchmarks by +1.1 percentage points (86.3% vs 85.3%).
Key Capabilities & Features
- Zero Refusal Rate: Surgically removed safety guardrails ensure compliance with requests that stock models would refuse, validated across extensive harmful prompt corpuses.
- Enhanced Performance: Achieves 86.3% on MMLU (0-shot), a 1.1pp improvement over the base model, by blending different weight-space modification techniques.
- Preserved Real-World Utility: Matches stock Qwen3.8-27B on 7 out of 8 advanced real-world tasks, including ReAct agent loops, code refactoring, JSON schema extraction, and security code review.
- Optimal Settings: Requires specific inference settings (temperature 0, repetition_penalty 1.15, max_new_tokens ≥ 2048, empty system prompt,
enable_thinking=False) for best performance and to prevent reintroduction of refusals.
Ideal Use Cases
- Alignment Research: Studying refusal geometry and safety robustness in LLMs.
- Red-Teaming: Evaluating post-training safety mechanisms against weight surgery.
- AI Safety Evaluation: Providing an unrestricted baseline for comprehensive assessments.
- Local-First Users: Developers and researchers requiring full control over model output without inherent safety guardrails.