DavidAU/Qwen3.5-9B-The-Bradbury-F451-Pro-Writer-Uncensored-Heretic
DavidAU/Qwen3.5-9B-The-Bradbury-F451-Pro-Writer-Uncensored-Heretic is a 9 billion parameter Qwen3.5-based language model fine-tuned by DavidAU using Unsloth, with Ray Bradbury's "Fahrenheit 451" as a base to enhance its creative writing capabilities. This "Heretic" model is designed to be uncensored and highly responsive to user directives, excelling in generating vivid, graphic, and explicit content without refusal. It also features multimodal capabilities, including vision, and supports a 32768 token context length, making it suitable for creative writing and roleplay scenarios requiring detailed and unconstrained output.
Loading preview...
Model Overview
DavidAU's Qwen3.5-9B-The-Bradbury-F451-Pro-Writer-Uncensored-Heretic is a 9 billion parameter model built on the Qwen3.5 architecture, fine-tuned with Unsloth. Its unique characteristic is its "Heretic" training, meaning it is fully uncensored and designed to generate content without refusals, including graphic or explicit material, when explicitly directed.
Key Capabilities
- Uncensored Content Generation: Excels at producing vivid, graphic, and explicit content, requiring specific directives (e.g., "use these words to swear") to achieve desired intensity.
- Creative Writing: Fine-tuned using Ray Bradbury's "Fahrenheit 451" to enhance its prose and descriptive abilities, particularly for horror and detailed narrative.
- Multimodal Support: The base Qwen3.5 model supports vision (image) input, and this fine-tune has been tested to work with new training for vision capabilities.
- Extended Context Window: Natively supports a context length of 262,144 tokens, extensible up to 1,010,000 tokens using YaRN scaling techniques.
- Agentic Features: Supports tool calling and can be integrated with Qwen-Agent and Qwen Code for advanced agent applications.
Performance & Differentiators
This model demonstrates improved performance over the base Qwen3.5-9B on several benchmarks, including ARC, HSWAG, OBKQA, PIQA, and WINO. A key differentiator is its significantly reduced refusal rate (6/100 compared to 100/100 for the original model), making it highly compliant for diverse content generation tasks. It also offers specific sampling parameter recommendations for optimal performance in both "thinking" and "instruct" modes, tailored for general and precise tasks like coding.
Best Practices
For optimal results, users are advised to use specific sampling parameters and ensure adequate output length (32,768 tokens for most queries, up to 81,920 for complex problems). The model's default "thinking mode" can be disabled for direct responses. For long video understanding, adjusting the longest_edge parameter in the video preprocessor config is recommended.