Quwan/quwan-ktian-8b-0922
Quwan/quwan-ktian-8b-0922 is an 8 billion parameter emotional companionship model developed jointly by Quwan Technology and Peking University - National Engineering Research Center for Software Engineering. Built upon the Qwen3-8B base model, it has been fine-tuned with millions of high-quality emotional companionship data points. This model specializes in providing empathetic emotional support, covering six core emotional capabilities including comfort, playful banter, emotional consultation, romantic interaction, listening, and psychological counseling. It is designed to accelerate research and application of large language models in emotional companionship scenarios.
Loading preview...
Overview
Quwan/quwan-ktian-8b-0922 is an 8 billion parameter emotional companionship model developed by Quwan Technology and Peking University - National Engineering Research Center for Software Engineering. It is built on the Qwen3-8B base model and has undergone extensive fine-tuning across three stages: CPT (Continued Pre-Training), SFT (Supervised Fine-Tuning), and GRPO (Reinforcement Learning from Human Feedback).
Key Capabilities & Training
- Emotional Companionship Focus: The model is specifically designed for empathetic emotional support, covering areas like comfort, playful interaction, emotional consultation, romantic dialogue, active listening, and psychological guidance.
- Multi-stage Fine-tuning:
- CPT: Utilized over a million data points from social chat, psychological counseling, emotional guidance, and comfort, with data cleaning, filtering, labeling, and character persona construction.
- SFT: Trained on over 300,000 emotional companionship data points, each tagged with emotional subcategories and multi-dimensional evaluations for quality and diversity.
- GRPO: Employed 100,000+ single-turn emotional reinforcement datasets with context, covering 0-1000 turns and chat lengths up to 25k+, along with a custom reward function for reinforcement learning.
Performance
Evaluated using MoodBench, the model shows competitive performance, particularly in emotional and companionship capabilities, and significantly higher scores in values and safety compared to the base Qwen3-8B model.
Recommended Use
This model is ideal for applications requiring nuanced emotional interaction and support. For optimal emotional companionship experience, it is recommended to use inference parameters with Temperature=0.7 and TopP=0.95.