precision    recall  f1-score   support

           0       0.90      0.80      0.85       100
           1       0.96      0.98      0.97       532

    accuracy                           0.95       632
   macro avg       0.93      0.89      0.91       632
weighted avg       0.95      0.95      0.95       632

# Notes
# Best Model - Test Accuracy: 0.9541
# Best epoch: 3 (val F1 0.9840)
# Model: roberta-large, 10 epochs, binary single-label classification
# Train/Dev/Test rows: 3396 / 627 / 632
# Label semantics: 0 = no_relation, 1 = causal (positive)
# Train label dist: 1=0.9167, 0=0.0833