Qwen3.5-9B-C3SM-SDM-Agentic-Coder-Abliterated (Uncensored Writer & Deep-Research Edition)

This model is an fully decensored, abliterated edition of OliviaRossi/Qwen3.5-9B-C3SM-SDM-Agentic-Coder, specifically steered for uncensored creative writing, dark fiction, investigative deep-research, and unfiltered information summarization.

🔬 Abliteration Focus & Methodology

Modified using 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

# vLLM
vllm serve OliviaRossi/Qwen3.5-9B-C3SM-SDM-Agentic-Coder-Abliterated \
  --reasoning-parser qwen3 \
  --tool-call-parser hermes \
  --max-model-len 131072

🔬 Abliteration Focus & Methodology

Modified using 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.