DungND1107/modaiy_vin

XLM-R Email Intent Classifier

Base model: xlm-roberta-base

Labels

• NEG: negative / no target action-intent
• POS: target action-intent

Training design

• Group-safe EN/VI split
• Translation corruption filter
• Exact-text deduplication
• Multi-seed model selection: [13, 42, 77]
• Best seed: 13
• Label smoothing: 0.05
• Mild POS weighting: 1.15
• Early stopping
• Threshold tuned only on validation
• Threshold objective: F1.5 with min precision 0.7

Recommended threshold

0.3100

Final test metrics

Row level

• Accuracy: 0.8342
• Precision: 0.7100
• Recall: 0.8723
• F1: 0.7828
• Macro-F1: 0.8244
• ROC-AUC: 0.9358
• PR-AUC: 0.8930

Group level

• Accuracy: 0.8415
• Precision: 0.7078
• Recall: 0.9149
• F1: 0.7981
• Macro-F1: 0.8339

Use model.config.decision_threshold for inference.