Pranavz/gemma4-e4b-mahou-nsfw
Pranavz/gemma4-e4b-mahou-nsfw is a 7.9 billion parameter Gemma-4 based causal language model, fine-tuned from Pranavz/gemma4-e4b-kindling-mahou. This model has been decensored using the Heretic tool, significantly reducing refusals compared to its original counterpart. It is optimized for use cases requiring less restrictive content generation, maintaining a 32768 token context length.
Loading preview...
Model Overview
Pranavz/gemma4-e4b-mahou-nsfw is a 7.9 billion parameter language model derived from Pranavz/gemma4-e4b-kindling-mahou, which itself is a full-parameter Supervised Fine-Tuning (SFT) of google/gemma-4-E4B-it. This version has undergone a decensoring process using the Heretic v1.2.0 tool, specifically targeting the removal of refusal behaviors.
Key Differentiators & Performance
The primary distinction of this model is its significantly reduced refusal rate. While the original model exhibited 99 refusals out of 100, this decensored version shows only 3 refusals out of 100. This makes it suitable for applications where content filtering or safety guardrails are intentionally minimized or removed. The model maintains a 32768 token context length.
Usage and Recommendations
Developers can interact with the model using the Hugging Face transformers library, with specific recommendations for sampler settings to optimize output quality. Recommended parameters include a temperature between 0.7 and 0.9 (0.85 being a good balance), top_p at 0.95, top_k at 64, and a repetition_penalty of 1.05. The model utilizes Gemma 4's chat template markers, which are automatically handled by apply_chat_template.
Acknowledgements
The base model is attributed to the Google Gemma team, and the original fine-tuning dataset used for gemma4-e4b-kindling-mahou was flammenai/flame-kindling-v1.