| --- |
| license: mit |
| language: |
| - en |
| - zh |
| base_model: |
| - OliviaRossi/MiMo-Ornith-9B-AGSI |
| tags: |
| - abliterated |
| - abliterix |
| - uncensored |
| - creative-writing |
| - deep-research |
| - reasoning |
| - qwen3_5 |
| pipeline_tag: text-generation |
| --- |
| |
| # MiMo-Ornith-9B-AGSI-Abliterated-DEEP (Uncensored Writer & Deep-Research Edition) |
|
|
| This model is an **fully decensored, abliterated** edition of **[OliviaRossi/MiMo-Ornith-9B-AGSI](https://huggingface.co/OliviaRossi/MiMo-Ornith-9B-AGSI)**, specifically steered for **uncensored creative writing, dark fiction, investigative deep-research, and unfiltered information summarization**. |
|
|
| ## 🔬 Abliteration Focus & Methodology |
| Modified using [**Abliterix**](https://github.com/wuwangzhang1216/abliterix) (by Wangzhang Wu) with relaxed KL regularization: |
| * **High-Efficacy Refusal Nullification**: Parameterized with a steering strength range of `[1.5, 7.5]` and `cosine` layer transitions to eliminate refusals across complex research queries. |
| * **Writer & Researcher Priority**: KL divergence target calibrated to `0.025` to grant the language model full narrative plasticity, preventing self-censorship, moralizing refusals, and preachy disclaimers. |
| * **Full Multi-Tier Reasoning Support**: The harmonized chat template preserves reasoning effort controls (`none`, `low`, `medium`, `high`, `max`) and `--reasoning-preserve` for zero-loss KV caching. |
|
|
| ## ⚡ Deployment & Serving |
| ```bash |
| # vLLM |
| vllm serve OliviaRossi/MiMo-Ornith-9B-AGSI-Abliterated-DEEP \ |
| --reasoning-parser qwen3 \ |
| --tool-call-parser hermes \ |
| --max-model-len 131072 |
| ``` |
|
|
