sulpikar2/Qwen3.5-9B-Uncensored-cyber-v3
sulpikar2/Qwen3.5-9B-Uncensored-cyber-v3 is a 9 billion parameter Qwen3.5-based causal language model, developed by sulpikar2, with a 32768 token context length. This model is a decensored iteration of sulpikar2/Qwen3.5-9B-Uncensored-cyber-v2, created using the Heretic tool. It demonstrates reduced refusal rates compared to its predecessor, making it suitable for applications requiring less restrictive content generation.
Loading preview...
Model Overview
sulpikar2/Qwen3.5-9B-Uncensored-cyber-v3 is a 9 billion parameter language model, developed by sulpikar2, building upon the Qwen3.5 architecture. This version is a decensored variant of the sulpikar2/Qwen3.5-9B-Uncensored-cyber-v2 model, achieved through the application of the Heretic v1.2.0 tool. It maintains a substantial context length of 32768 tokens.
Key Characteristics
- Decensored Nature: This model has undergone a decensoring process, resulting in a significantly lower refusal rate compared to its direct predecessor. Benchmarking shows a reduction from 23/100 refusals in the original model to 8/100 in this version.
- Reproducibility: The decensoring process is reproducible, with detailed instructions available in the
reproduce/README.mddirectory. - Training Efficiency: The base
Qwen3.5model was fine-tuned using Unsloth and Huggingface's TRL library, enabling 2x faster training.
Abliteration Parameters
Specific abliteration parameters were applied per layer to achieve the decensored state, including adjustments to attn.o_proj and mlp.down_proj weights, with values such as attn.o_proj.max_weight at 1.10 and mlp.down_proj.min_weight at 0.18.
Use Cases
This model is particularly suited for applications where a less restrictive content generation policy is desired, offering a lower likelihood of refusing prompts compared to its more censored counterparts.