holi-lab/ArcANE-32B-SFT
ArcANE-32B-SFT by holi-lab is a 32 billion parameter Qwen3-based model fine-tuned with LoRA for advanced English role-playing. It specializes in generating character responses that accurately reflect a character's evolving behavioral state and value changes across narrative phases, rather than a static persona. This model is particularly suited for research into point-in-time character role-play and responses conditioned on chapter-truncated Character Arcs, achieving an overall score of 58.4 in specialized evaluation metrics.
Loading preview...
ArcANE-32B-SFT: Dynamic Character Role-Playing
ArcANE-32B-SFT is a 32 billion parameter model developed by holi-lab, based on the Qwen3-32B architecture and fine-tuned using LoRA. This model is specifically designed for generating sophisticated English role-playing responses that capture a character's evolving behavioral and value states throughout a narrative, rather than a fixed persona. It was accepted to the EMNLP 2026 Main Conference.
Key Capabilities & Features
- Point-in-Time Role-Play: Generates character responses conditioned on their specific behavioral state at a given point in a narrative.
- Character Arc Conditioning: Utilizes chapter-truncated Character Arcs to inform responses, ensuring fidelity to the character's development.
- Behavioral & Value Change Reflection: Excels at reflecting how characters change across different narrative phases.
- LoRA Fine-Tuning: Trained on the supervised split of the ArcANE dataset, derived from 12 training novels covering 55 characters and 339 character axes.
- Performance: Achieves an overall score of 58.4 in specialized evaluation metrics (APF, RPF, RAE, PTF) with Arc context, outperforming its non-Arc context counterpart (53.7) and the base Qwen3-32B (50.1).
Intended Use Cases
- Research into dynamic character role-playing and agent behavior.
- Developing applications requiring character responses conditioned on evolving narrative contexts.
- Exploration of behavioral and value changes in characters across story phases.
Users should supply the relevant Character Arc only up to the queried chapter, ensuring future phases are not exposed to the model for faithful point-in-time conditioning.