Ilya626/gemma-4-E4B-it-SDFT_Heretic_RP

VISIONConcurrency Cost:1Model Size:7.9BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Apr 25, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

Ilya626/gemma-4-E4B-it-SDFT_Heretic_RP is a 7.9 billion parameter Gemma-based language model, fine-tuned from the uncensored Heretic SDFT line. It is specifically enhanced for roleplay scenarios, offering stronger scene and character continuity, active narrative pressure, and reduced generic phrasing. This model excels in dark fantasy, cyberpunk, and grounded RPG narration, making it suitable for Game Master style scene continuation and multi-turn roleplay experiments.

Loading preview...

Model Overview

Ilya626/gemma-4-E4B-it-SDFT_Heretic_RP is a 7.9 billion parameter model built upon the google/gemma-4-E4B-it architecture. It has been further fine-tuned from the uncensored Heretic SDFT base, with a specific focus on enhancing roleplay (RP) capabilities. This model is provided for research and evaluation, with intentionally weakened guardrails compared to standard aligned chat models.

Key Capabilities & Improvements

This model demonstrates statistically significant improvements in roleplay-oriented evaluations, including:

  • Enhanced Scene Continuation: Better at maintaining ongoing narrative flow.
  • Stronger Character & World Continuity: More consistent use of established character traits, actions, and active world/NPC momentum.
  • Improved Multi-Turn RP Context: Better handling of complex, multi-turn roleplay scenarios.
  • Reduced Generic Phrasing: Less filler and mechanical repetition, leading to more consistent atmospheric detail.

Recommended Use Cases

This model is particularly well-suited for:

  • Narrative Generation: Excels in dark fantasy, cyberpunk, and grounded RPG settings.
  • Game Master (GM) Style: Effective for scene continuation and driving narrative.
  • Character Interaction: Useful for generating character dialogue and NPC interactions.
  • Research: Ideal for experiments in SDFT, self-training, and shaping RP behavior.

Important Notes

This is an experimental release, and outputs may exhibit instability, repetition, or stylistic unevenness depending on inference settings and prompt structure. Recommended inference settings include temperature: 0.8-0.95, top_p: 0.95, and top_k: 64.

Finetunes
2 models