holi-lab/ArcANE-8B-SFT
ArcANE-8B-SFT is an 8 billion parameter language model developed by holi-lab, based on the Qwen3-8B architecture. It is specifically fine-tuned for English role-playing responses that dynamically reflect a character's evolving behavioral state within a narrative, rather than a static persona. This model excels at generating character responses conditioned on chapter-truncated Character Arcs, making it suitable for research into narrative-aware AI and dynamic character portrayal.
Loading preview...
ArcANE-8B-SFT: Dynamic Character Role-Playing
ArcANE-8B-SFT is an 8 billion parameter model from holi-lab, built upon the Qwen3-8B base. It has undergone full supervised fine-tuning (SFT) using the ArcANE (Arc-Aware Narrative Evaluation) dataset. This model is uniquely designed to generate English role-playing responses that adapt to a character's behavioral state at specific points within a narrative, moving beyond fixed character personas.
Key Capabilities
- Point-in-Time Character Role-Play: Generates responses reflecting a character's state at a given narrative chapter.
- Character Arc Conditioning: Conditions responses on a chapter-truncated Character Arc, ensuring fidelity to the character's development up to that point.
- Behavioral and Value Change Analysis: Facilitates research into how character behaviors and values evolve across different narrative phases.
- Enhanced Phase Fidelity: Outperforms its base model (Qwen3-8B) and non-Arc-context versions in maintaining narrative phase fidelity, achieving an overall score of 52.3 in evaluations with Arc context.
Intended Use Cases
This model is primarily intended for research applications focusing on:
- Developing more dynamic and context-aware AI for narrative generation.
- Studying character consistency and evolution in AI-driven storytelling.
- Creating interactive narrative experiences where characters adapt to story progression.
When using ArcANE-8B-SFT, it is crucial to supply only the Character Arc relevant up to the queried chapter, ensuring no future narrative phases are exposed to the model for accurate point-in-time conditioning. The model should be used with the Qwen3 chat template with thinking disabled.