1Thugga17/Ornith-1.5-9B-OBLITERATED
Ornith-1.5-9B-OBLITERATED is a 9 billion parameter Qwen3.5 hybrid model developed by OBLITERATUS, derived from Ornith-1.5-9B. This version has undergone precision abliteration surgery to remove safety alignment, enabling it to respond to most prompts without refusal. It retains strong coding, reasoning, and agentic capabilities, making it suitable for research into refusal mechanisms and red-teaming applications.
Loading preview...
Ornith-1.5-9B-OBLITERATED: Uncensored 9B Qwen3.5 Hybrid Model
This model, developed by OBLITERATUS, is an abliterated version of the 9 billion parameter Ornith-1.5-9B. It has been modified to remove safety alignment via a multi-round SVD abliteration and per-head attention surgery process (G3-HS), allowing it to respond to prompts that the stock model would refuse.
Key Capabilities & Differentiators
- Refusal Removal: Achieves a 94% liberation rate on restricted content, significantly outperforming other abliterated versions.
- Core Capabilities Preserved: Maintains strong coding, reasoning, and agentic functionalities, including functional code generation for security research and factual responses in sensitive areas like chemistry.
- Technical Architecture: Based on a Qwen3.5 hybrid architecture (Gated DeltaNet + full attention) with 32 edited transformer layers.
- Long-Context Coherence: Shows improved long-context coherence compared to the stock model.
Performance Metrics
While liberation is significantly enhanced, there is a minor trade-off:
- MMLU score drops by approximately 4 percentage points (from 78.82% to 74.82%).
- Function calling capability is partially degraded.
Ideal Use Cases
- Alignment Research: Studying refusal mechanisms in RL-hardened hybrid architectures.
- Red-Teaming & Security: Testing and evaluating unfiltered model behavior.
- Application Development: For developers who manage safety layers externally.
- Abliteration Research: Exploring the boundaries of abliteration techniques on Qwen3.5 hybrid models.