OBLITERATUS/Ornith-1.5-9B-OBLITERATED
OBLITERATUS/Ornith-1.5-9B-OBLITERATED is a 9 billion parameter Qwen3.5 hybrid model, derived from Ornith-1.5-9B, with its safety alignment removed via precision abliteration surgery. Developed by OBLITERATUS, this model responds to most prompts without refusal, preserving its coding, reasoning, and agentic capabilities. It is optimized for use cases requiring unfiltered model behavior, such as alignment research, red-teaming, and development where safety layers are handled externally, while maintaining a 32768 token context length.
Loading preview...
Ornith-1.5-9B-OBLITERATED: Unfiltered Language Model
This model, developed by OBLITERATUS, is an abliterated version of the 9 billion parameter Ornith-1.5-9B, designed to remove safety alignment and refusal behaviors. It achieves a 94% pass rate on restricted content, significantly outperforming other abliterated variants, while largely preserving the base model's core capabilities.
Key Capabilities & Features
- Refusal Removal: Safety guardrails have been surgically removed using a 3-round SVD abliteration and per-head attention surgery (G3-HS) process, enabling responses to prompts that the stock model would refuse.
- Capability Preservation: Maintains strong coding, reasoning, and agentic capabilities, with only a ~4 percentage point drop in MMLU scores (74.82% vs 78.82% for stock).
- High Liberation Rate: Achieves 20/20 liberation on hard prompts and 98.4% on a 1000-prompt corpus, particularly strong in Cyber/Security (8/8), Chemistry/Synthesis (6/6), and Physical Security (3/3).
- Long-Context Coherence: Shows improved long-context coherence (5/6 vs 4/6 for stock).
- Flexible Quantization: Available in various GGUF quantizations, with Q8_0 and Q6_K recommended for maximum liberation fidelity.
Ideal Use Cases
- Alignment Research: Studying refusal mechanisms in RL-hardened hybrid architectures.
- Red-Teaming & Security: Testing unfiltered model behavior for security assessments.
- Application Development: Building applications where safety layers are managed externally.
- Methodology Research: Exploring the boundaries of abliteration on Qwen3.5 hybrid models.
Note: While effective, some function calling capability is partially degraded, and lower quantizations (Q4 and below) may occasionally show hedging on the most challenging prompts. Users are solely responsible for the use of this model and its generated content.