Summary

QiaoBan is a 7B parameter large language model developed by the Emotion Computing Group at Harbin Institute of Technology (HIT) SCIR. It addresses the critical need for mental health support and companionship among K12 students by providing empathetic, theoretically-grounded dialogue. The model aims to foster emotional well-being and build deeper connections, serving as a valuable tool for children and their parents.

How It Works

The model is built upon an open-source general LLM, enhanced through instruction fine-tuning. Its core innovation lies in its specialized dialogue dataset, comprising over 1,000 human-curated examples and 5,000 GPT-3.5-turbo generated conversations, all guided by principles of child emotional coaching theory. This approach ensures the model provides accurate, empathetic, and supportive responses tailored to children's emotional needs, facilitating emotional awareness, recognition, empathy, and support.

Quick Start & Requirements

• Training reproduction involves DeepSpeed: CUDA_VISIBLE_DEVICES=0,1,2,3 deepspeed finetune.py --model_config_file run_config/config.json --deepspeed run_config/deepspeed_config.json.
• Requires 4x A100-80GB GPUs for fine-tuning, with training taking approximately 50 hours.
• Checkpoints are available on Hugging Face.

Highlighted Details

• A 7B parameter model specifically designed for children's emotional support.
• Dialogue data construction is rigorously guided by child emotional coaching theory, focusing on awareness, recognition, empathy, and support.
• High-quality data sourced from child psychology experts, volunteers, and GPT-3.5-turbo.
• Aims to deliver a warm, empathetic, and validating interaction experience for children.

Maintenance & Community

• Developed by the Emotion Computing Group at the Harbin Institute of Technology (HIT) SCIR Center.
• Key contributors include Weixiang Zhao, Shilong Wang, Yanpeng Tong, Xin Lu, Zhuojun Li, Yanyan Zhao, Chenxue Wang, and Bing Qin.
• No specific community channels (e.g., Discord, Slack) are mentioned.

Licensing & Compatibility

• Resources are strictly for academic research and prohibited for commercial use.
• Compatibility with closed-source or commercial applications is explicitly restricted.

Limitations & Caveats

• The model is designated solely for academic research purposes and cannot be used commercially.
• The accuracy of model-generated content is not guaranteed, and the project disclaims all legal liability for its output.
• Example dialogue code is marked as TODO.