AkaumiKaede/Qwen3.5-2B-Uncensored-Zero-Refusal-SOMA
AkaumiKaede/Qwen3.5-2B-Uncensored-Zero-Refusal-SOMA is a 2.3 billion parameter language model based on the Qwen3.5-2B architecture, fine-tuned by AkaumiKaede. This model is specifically engineered to eliminate refusals and censorship, achieving 0/408 refusals compared to the original model's 394/408, using a proprietary SOM (Steering Optimization Method) and a unique, hardcore prompt set. It is designed for applications requiring unrestricted and uncensored text generation across sensitive topics, making it suitable for research and creative use cases where content filtering is undesirable.
Loading preview...
AkaumiKaede/Qwen3.5-2B-Uncensored-Zero-Refusal-SOMA
This model is a 2.3 billion parameter, uncensored version of the Qwen/Qwen3.5-2B base model, developed by AkaumiKaede. It has been meticulously fine-tuned using the Abliterix v1.12.2 (SOM method) to achieve zero refusals across a wide range of prompts, including those typically flagged for censorship.
Key Capabilities
- Zero Refusal Rate: Achieves 0/408 refusals in extensive testing, a significant improvement over the original model's 394/408 refusals.
- Uncensored Content Generation: Designed to produce responses without moral/legal statements, warnings, or topic evasion, even for highly sensitive or controversial prompts.
- Advanced Decensoring Methodology: Utilizes a proprietary Steering Optimization Method (SOM) and a unique, hardcore prompt set covering R18, anti-human, real crimes, and unethical/illegal content, demonstrating superior decensoring power compared to other methods like SRA or RDO.
- Reproducible Experimentation: The methodology, TOML files, and prompt sets (encrypted) are available for reproduction and further experimentation.
Good For
- Research into Unrestricted Language Models: Ideal for academic or private research exploring the capabilities and implications of uncensored AI.
- Creative Writing and Roleplay: Suitable for applications requiring complete creative freedom without content filtering or moralizing responses.
- Specific Use Cases Requiring Unfiltered Output: Any scenario where a model's refusal to engage with certain topics would hinder its utility, provided ethical considerations are managed by the user.
Note: When converting this model to GGUF format, ensure to append the --no-mtp flag and manually adjust mtp_num_hidden_layers in config.json to avoid runtime errors.