YuYu1015/YuYu1015-Ornith-1.0-9B-abliterated
YuYu1015/YuYu1015-Ornith-1.0-9B-abliterated is a 9 billion parameter Qwen3.5-architecture model developed by YuYu1015. This variant is an 'abliterated' (uncensored) version of deepreinforce-ai/Ornith-1.0-9B, specifically designed to remove refusal behavior while fully preserving its original reasoning capabilities. It excels at complex reasoning tasks, maintaining high accuracy on benchmarks like GSM8K, and supports both English and Chinese languages.
Loading preview...
Model Overview
YuYu1015-Ornith-1.0-9B-abliterated is an uncensored variant of the 9 billion parameter Qwen3.5-architecture reasoning model, deepreinforce-ai/Ornith-1.0-9B. Developed by YuYu1015, this model focuses on removing refusal behavior and censorship without compromising the base model's core reasoning and thinking abilities. It utilizes a Qwen3.5 9B dense architecture, incorporating GatedDeltaNet (linear attention) and a 3:1 full-attention hybrid.
Key Capabilities & Differentiators
- Censorship Removal: Achieves a hard refusal rate of less than 1% (down from 99.5% in the base model) by being a weights-only modification with zero additional training, ensuring behavioral fidelity.
- Reasoning Preservation: Maintains the base model's strong reasoning performance, with GSM8K accuracy at 85%, virtually unchanged from the original 86.7%.
- Multilingual Support: Operates in both English and Chinese.
- Thinking Mode: Supports a dedicated thinking mode, emitting
<think>…</think>tokens, characteristic of its reasoning-focused design.
Important Usage Notes
This model is highly sensitive to sampling parameters. It requires a repeat-penalty of 1.05 to function correctly; deviations can lead to severe thinking loops or truncated responses. A DPO-fine-tuned variant (-dpo) is also available for further moralizing reduction, though it may alter the model's behavior more aggressively.