DavidAU/Qwen3.5-27B-Polar-Rev1-Uncensored-Heretic
DavidAU/Qwen3.5-27B-Polar-Rev1-Uncensored-Heretic is a 27 billion parameter Qwen 3.5 model fine-tuned by DavidAU using the advanced "Polaris Dataset" and a unique "Rev1" training method. This model demonstrates improved performance across 7 key benchmarks compared to the base Qwen 3.5 27B model, despite using a small, non-reasoning dataset of only 1017 samples. It is also uncensored via the HERETIC method, making it suitable for generating content without refusals.
Loading preview...
Model Overview
DavidAU/Qwen3.5-27B-Polar-Rev1-Uncensored-Heretic is a 27 billion parameter Qwen 3.5 model, fine-tuned by DavidAU using a novel "Rev1" training method and the "Polaris Dataset" from TeichAI. This model showcases significant performance improvements, with benchmarks indicating gains across 7 out of 7 key metrics when compared to the untuned Qwen 3.5 27B model. Notably, these improvements were achieved with a remarkably small dataset of only 1017 non-reasoning samples.
Key Capabilities
- Enhanced Performance: Demonstrates improved scores across multiple benchmarks (arc, arc/e, boolq, hswag, obkqa, piqa, wino) compared to the base Qwen 3.5 27B-Instruct model.
- Uncensored Output: Utilizes the "HERETIC" method, ensuring the model generates content without refusals, even for sensitive or explicit requests. It requires specific directives (e.g., using slang or explicit terms) to generate content at expected graphic levels.
- Vision Capabilities: The underlying Qwen 3.5 architecture supports vision (images), which has been tested with new training.
- Efficient Training: Achieved substantial gains with a very small dataset, highlighting the effectiveness of the "Rev1" training method.
Good For
- Unrestricted Content Generation: Ideal for use cases requiring the generation of content without built-in refusals or censorship, provided appropriate prompting is used.
- Creative Writing & Roleplay: The uncensored nature and ability to be 'pushed' with directives make it suitable for generating vivid, intense, or graphic narratives, as demonstrated by the example generation.
- Exploration of Fine-tuning Techniques: Offers insights into the impact of specific training methods and small, high-quality datasets on model performance.