DavidAU/Qwen3.5-9B-Polar-Rev1-Uncensored-Heretic
DavidAU/Qwen3.5-9B-Polar-Rev1-Uncensored-Heretic is a 9 billion parameter Qwen3.5-based instruction-tuned language model with a 32768 token context length. It was fine-tuned by DavidAU using the 'Polaris Dataset' and a unique 'Rev1' training method, demonstrating improved performance across various benchmarks compared to its base model and even larger 27B models. This model is also uncensored via 'HERETIC' prior to training, making it suitable for use cases requiring unfiltered content generation.
Loading preview...
Overview
DavidAU/Qwen3.5-9B-Polar-Rev1-Uncensored-Heretic is a 9 billion parameter instruction-tuned model built on the Qwen3.5 architecture. It features a 32768 token context length and was fine-tuned by DavidAU using a novel "Rev1" training method and the "Polaris Dataset." This approach significantly enhances its performance, with the 9B 'Rev1' version outperforming the non-trained 27B Qwen3.5 model in 5 out of 7 benchmarks, and nearly matching it in the remaining two. A key characteristic is its uncensored nature, achieved through a "HERETIC" process applied before training, ensuring it generates content without refusal.
Key Capabilities
- Enhanced Performance: Demonstrates superior benchmark scores (e.g., arc, boolq, obkqa, piqa, wino) compared to its base model and larger counterparts, despite being trained on a small, non-reasoning dataset of only 1017 samples.
- Uncensored Output: Designed to generate content without refusals, offering full creative freedom, though it may require explicit direction for highly graphic or explicit content.
- Multimodal Support: Supports vision inputs, with image and video processing capabilities, and is compatible with various inference frameworks like SGLang, vLLM, and Hugging Face Transformers.
- Long Context Handling: Natively supports up to 262,144 tokens and can be extended to 1,010,000 tokens using RoPE scaling techniques like YaRN.
Good for
- Unrestricted Content Generation: Ideal for applications where content filtering or refusals are undesirable, such as creative writing, roleplay, or specific technical simulations.
- Performance-Critical Tasks: Suitable for tasks requiring strong performance in reasoning and general language understanding, as indicated by its benchmark improvements.
- Multimodal Applications: Can be used in scenarios involving image and video input, leveraging its unified vision-language foundation.
- Agentic Workflows: Recommended for building agent applications, with specific optimizations for tool calling and integration with frameworks like Qwen-Agent and Qwen Code.