holi-lab/ArcANE-32B-RLVR

TEXT GENERATIONPricing:Input $0.408 / Cached $0.0816 / Output $1.972Concurrent Unit Cost:2Model Size:32BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 5, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

holi-lab/ArcANE-32B-RLVR is a 32 billion parameter Qwen3-based model developed by holi-lab, specifically trained with Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and Group Relative Policy Optimization (GRPO for Reinforcement Learning with Verifiable Rewards). This model is optimized for point-in-time character role-play and generating character responses conditioned on chapter-truncated Character Arcs, using a rubric-based reward system for narrative phase fidelity. It excels in research applications requiring nuanced character interaction within narrative contexts, demonstrating an 8.0 point improvement in Overall score over its DPO predecessor.

Loading preview...

ArcANE-32B-RLVR: Narrative Role-Playing with Verifiable Rewards

ArcANE-32B-RLVR is a 32 billion parameter model built upon the Qwen/Qwen3-32B architecture, developed by holi-lab. It has undergone a multi-stage training process, including Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and a novel Reinforcement Learning with Verifiable Rewards (RLVR) phase utilizing Group Relative Policy Optimization (GRPO). This model continues from ArcANE-32B-DPO, employing the ArcANE per-phase evaluation rubric as its reward signal.

Key Capabilities

  • Point-in-time character role-play: Designed to maintain character consistency and responses at specific narrative junctures.
  • Character responses conditioned on Character Arcs: Generates dialogue and actions based on a chapter-truncated Character Arc, ensuring fidelity to the character's development up to a given point.
  • Rubric-based reinforcement learning: Leverages a unique reward system for narrative phase fidelity, improving the quality and relevance of generated content.
  • Improved performance: Achieves an "Overall" score of 68.2, an 8.0 point improvement over ArcANE-32B-DPO in shared-record evaluations.

Good for

  • Research into narrative AI: Ideal for exploring advanced techniques in character consistency and story generation.
  • Developing interactive narrative experiences: Useful for creating dynamic character interactions that evolve with the story.
  • Studying reinforcement learning for creative tasks: Provides a strong baseline for research into reward modeling for complex, subjective outputs like narrative.

For reproducibility, the training recipe is available in the ArcANE repository.

Finetunes
1 models