johnnycpu/chinese-mentalbert-crisis-detection

微光導航 LightGuide — Task E — Crisis Emotion Detection

Chinese-MentalBERT fine-tuned for 3-class mental health emotion classification (POSITIVE / NEGATIVE / CRISIS) with Focal Loss (gamma=2, adaptive alpha). Specialized for detecting implicit crisis signals in elderly welfare service conversations. Achieves 100% recall on implicit crisis test cases. Part of the 微光導航 (LightGuide) project.

Model Details

• Base Model: zwzzz/Chinese-MentalBERT
• Fine-tuning Task: Task E — Crisis Emotion Detection
• Project: 微光導航 (LightGuide) — AI-powered elderly welfare navigation system
• Competition: 大專生資訊服務創新競賽 (Taiwan University Information Service Innovation Competition)
• License: Apache 2.0

Usage

from transformers import AutoTokenizer, AutoModel
from peft import PeftModel  # if LoRA adapter

# Load from Hub
model = AutoModel.from_pretrained("johnnycpu/chinese-mentalbert-crisis-detection")

Training Details

• Dataset: Custom Traditional Chinese welfare dialogue dataset
• Training Hardware: NVIDIA A100 (Google Colab)
• Framework: HuggingFace Transformers + PEFT

Citation

If you use this model, please cite the 微光導航 project.