XLM-RoBERTa-base fine-tuned for 5-class emotion detection on English, Filipino, and Taglish diary entries.

Classes: angry, anxious, happy, neutral, sad

Architecture: XLM-RoBERTa-base + custom classification head (Dropout → Linear → LayerNorm → GELU → Dropout → Linear)

Training: Focal Loss + Label Smoothing, SWA, Differential LR, Keyword hybrid layer

See the DiariCore notebook for full inference code including calibration multipliers and keyword assist layer.