This directory contains an improved LLM-based data augmentation system specifically designed to address severe class imbalance in the CMCC-34 dataset. The system uses GLM-4 model to generate high-quality synthetic samples for minority classes.

# Install required packages
pip install -r requirements.txt

# Start the GLM-4 API server (runs on port 8001)
cd ../inference && python glm4v_server.py

You should keep the server running until the augmentation is finished.

cd ../aug && python run_augmentation.py              # Run full augmentation

# Generate comprehensive analysis and visualizations
python visualization.py

cd ../finetune && python finetune.py  ../aug/data THUDM/GLM-4-9B-0414  configs/balanced_qlora.yaml 

python evaluate.py --model-path ../finetune/output/cmcc34_qlora_balanced_system_prompt/checkpoint-5000

• GLM-4 Model: THUDM/GLM-4-9B-0414
• CMCC-34 Dataset: Telecom customer service intent classification
• Evaluation Metrics: Gini coefficient, F1-score, precision/recall
• Technical Paper: [LLM-based Data Augmentation for Imbalanced Text Classification]

💡 Pro Tip: Start with dry-run mode to verify configuration, then run full augmentation during off-peak hours due to computational requirements.