YuYu1015/YuYu1015-Ornith-1.0-9B-abliterated-dpo
YuYu1015-Ornith-1.0-9B-abliterated-dpo is a 9 billion parameter Qwen3.5-architecture reasoning model, developed by YuYu1015, that has undergone abliteration and DPO fine-tuning. This variant significantly reduces moralizing and refusal tendencies (to 31% and <1% respectively) while maintaining or slightly improving its reasoning ability, achieving 90.0% on GSM8K. It is designed for applications requiring uncensored responses and strong reasoning in both English and Chinese, with a context length inherited from its base model.
Loading preview...
Overview
This model, YuYu1015-Ornith-1.0-9B-abliterated-dpo, is a 9 billion parameter Qwen3.5-architecture reasoning model. It is an abliterated (uncensored) and DPO fine-tuned variant of deepreinforce-ai/Ornith-1.0-9B, developed by YuYu1015. The primary goal of this fine-tuning was to remove refusal behavior and substantially reduce moralizing/disclaimer tendencies, while preserving and even slightly improving its core reasoning capabilities.
Key Capabilities & Differentiators
- Reduced Censorship: Hard refusal rates are brought down to less than 1% (from 99.5%), and moralizing/disclaimer rates are reduced to 31% (from 99.5%).
- Enhanced Reasoning: Despite the behavioral modifications, the model's reasoning accuracy on GSM8K has slightly improved from 86.7% to 90.0%.
- Qwen3.5 Architecture: Utilizes a GatedDeltaNet (linear attention) + full-attention hybrid (3:1) structure.
- Thinking Mode Support: As a reasoning model, it supports and emits
<think>…</think>tokens. - Multilingual: Supports both English and Chinese languages.
Important Usage Notes
- This model requires specific sampling parameters, particularly a
repeat-penaltyof1.05. Deviating from this value can lead to severe thinking loops or truncated answers. - A pure abliterated variant (
-abliterated) is also available, offering slightly higher moralizing (38%) but with reasoning fully intact and closer to the base model's original behavior.
Safety Warning
As safety filtering has been removed, this model may generate sensitive or inappropriate content. Users are responsible for ensuring compliance with local laws and ethical standards.