llmfan46/Qwen3.5-9B-Nikusui-v1-Uncensored-Heretic-Native-MTP-Preserved
llmfan46/Qwen3.5-9B-Nikusui-v1-Uncensored-Heretic-Native-MTP-Preserved is a 9 billion parameter Qwen3.5-based causal language model developed by llmfan46, featuring a 32768-token context length. This model is a decensored version of extraltodeus/Qwen3.5-9B-Nikusui-v1, engineered using Heretic v1.4.0 and the Magnitude-Preserving Orthogonal Ablation (MPOA) method. It significantly reduces refusal rates by 86% (11/100) compared to the original model (96/100) while preserving quality with a low KL divergence of 0.0067. The model is optimized for use cases requiring less content refusal without substantial performance degradation.
Loading preview...
Model Overview
llmfan46/Qwen3.5-9B-Nikusui-v1-Uncensored-Heretic-Native-MTP-Preserved is a 9 billion parameter language model built upon the Qwen3.5-9B-Base architecture, offering a 32768-token context window. This model is a decensored variant of extraltodeus/Qwen3.5-9B-Nikusui-v1, created by llmfan46 using the Heretic v1.4.0 tool and a Magnitude-Preserving Orthogonal Ablation (MPOA) method.
Key Differentiators & Performance
This model's primary distinction is its significantly reduced refusal rate, achieving 11 refusals out of 100, an 86% reduction compared to the original model's 96 refusals out of 100. This decensoring process maintains model quality, as indicated by a low KL divergence of 0.0067 from the original. The ablation targeted specific components including attn.out_proj, mlp.down_proj, and attn.o_proj to achieve this behavior modification.
MMLU Benchmarks
The model demonstrates strong general knowledge capabilities, with an overall MMLU accuracy of 77.10%, closely matching the original model's 77.28%. This indicates that the decensoring process did not substantially degrade its performance on a wide range of academic and professional subjects.
Use Cases
This model is particularly suited for applications where a lower propensity for content refusal is desired, without a significant compromise in factual accuracy or general language understanding. It is designed for users who require a more permissive model output while retaining the core capabilities of the Qwen3.5-9B-Nikusui-v1 base.