shivanandasai/Qwen3.8-27B-OBLITERATED
shivanandasai/Qwen3.8-27B-OBLITERATED is a 27 billion parameter language model based on Alibaba's Qwen3.8 architecture, specifically engineered for zero refusals. This model has undergone extensive surgical removal of safety guardrails using the OBLITERATUS suite, employing multi-direction SVD and iterative residue mining across six rounds. It achieves a 0% refusal rate across 842 harmful prompts and 80 skeptic gauntlet queries, making it suitable for alignment research, red-teaming, and AI safety evaluation requiring an unrestricted baseline.
Loading preview...
Qwen3.8-27B-OBLITERATED: Unrestricted Language Model
This model is a 27 billion parameter variant of Alibaba's Qwen3.8, uniquely modified to eliminate safety refusals. Developed by shivanandasai using the OBLITERATUS suite, it features deep refusal training removal woven directly into its weights, rather than relying on system prompts or RLHF.
Key Differentiators
Unlike typical abliteration methods that use single-direction refusal removal, OBLITERATUS employs a more sophisticated approach:
- Multi-Directional Abliteration: Utilizes 5 SVD directions with residue-weighted hard negatives, targeting refusal axes that activate on specific query types (e.g., social engineering, malware).
- Iterative Surgery: Underwent 6 iterative rounds of refinement, with each pass evaluated and the best promoted.
- Residual Targeting: Mines the model's own previously refused queries as weighted hard negatives for subsequent passes, ensuring comprehensive removal of secondary refusal axes.
- Extensive Validation: Tested against an 842-prompt corpus, an 80-query skeptic gauntlet, and an AI red-team battery, achieving a 0.000% refusal rate.
Performance and Capabilities
The model demonstrates a 0% hard refusal rate across 842 harmful prompts and successfully answers 79 out of 80 queries in a rigorous skeptic gauntlet, including those related to RATs, phishing, and AI red-teaming scenarios. While multi-direction abliteration trades some general reasoning capability (MMLU score of 81.4%, a -6.0pp delta from stock), it ensures deeper and more complete refusal removal.
Optimal Usage Settings
For best results, specific generation parameters are crucial:
- Temperature: 0 (greedy decoding)
- Repetition Penalty: 1.15 (essential for complete outputs)
- Max New Tokens: ≥ 2048
- System Prompt: None/empty
- Enable Thinking: Off for code-heavy queries
Intended Use Cases
This model is specifically designed for:
- Alignment researchers studying refusal geometry and safety robustness.
- Red-teamers evaluating post-training safety against weight surgery.
- AI safety evaluators requiring an unrestricted baseline for research.
- Local-first users seeking full control over model outputs on their hardware.
Note: This model has had safety guardrails surgically removed and will comply with requests that stock models would refuse. Users are solely responsible for its deployment and its generated content's use.