Class Distillation with Mahalanobis Contrast: An Efficient Training Paradigm for Pragmatic Language Understanding Tasks

This provides the code for ClaD model training and evaluation on misogyny data, split into two main files:

File Structure

• ClaD_misogyny.py

  • Contains the model definition, training code, and logic for saving model parameters.
  • Key features:
    • Load pre-trained models (e.g., XLNet).
    • Build task-specific structures (supports adding extra layers, freezing certain layers, etc.).
    • Define loss functions (e.g., Mahalanobis mean loss) and perform training.
    • Save the trained model weights.

• evaluation_xlnet_misogyny.ipynb

  • Used for model evaluation and analysis.
  • Key features:
    • Load the trained model weights.
    • Run predictions on test data.
    • Compute various evaluation metrics (e.g., accuracy, F1 score, Mahalanobis distance).
    • Support result visualization for better performance analysis.