Pranavz/qwen-4b-2507-abil-mahou
Pranavz/qwen-4b-2507-abil-mahou is a 4 billion parameter Qwen3-based causal language model, developed by Pranavz, with a 32K context length. This model is a decensored version of Pranavz/qwen-4b-2507-rp-mahou, specifically modified to reduce refusals and enhance creative roleplay and character interaction capabilities. It is fine-tuned for vivid, action-asterisk style creative writing, making it suitable for applications requiring less restrictive content generation.
Loading preview...
Model Overview
Pranavz/qwen-4b-2507-abil-mahou is a 4 billion parameter language model based on the Qwen3 architecture, developed by Pranavz. It is a decensored variant of the Pranavz/qwen-4b-2507-rp-mahou model, created using the Heretic v1.2.0 tool. The primary goal of this modification is to significantly reduce content refusals, making it more permissive for creative generation.
Key Capabilities & Differentiators
- Decensored Output: Achieves a refusal rate of 5/100, a substantial reduction compared to the original model's 99/100, enabling broader creative expression.
- Creative Roleplay: Fine-tuned specifically for creative roleplay and character interaction, utilizing the
flammenai/flame-kindling-v1dataset. - Vivid Writing Style: Optimized for generating vivid, descriptive text, often employing asterisks for actions, consistent with roleplay conventions.
- Full-Sequence SFT: Trained using full-sequence Supervised Fine-Tuning (SFT) on the Qwen3-4B-Instruct-2507 base model.
- Context Length: Supports a context length of 32,768 tokens.
Use Cases & Limitations
This model is particularly well-suited for applications requiring creative writing, immersive roleplay, and character-driven narratives where a less restrictive content filter is desired. Users should note its specific training on a curated roleplay dataset, which imparts a particular tone and style. It is not safety-tuned beyond its base model and is English-only.
For optimal performance in roleplay scenarios, it is recommended to use specific sampler settings (e.g., temperature 0.7-0.85, repetition_penalty 1.05-1.15) and to pass enable_thinking=False to the chat template to prevent Chain-of-Thought (CoT) reasoning.