Kaaeun/Labeling

KcELECTRA-base-finetuned-suicide-comments

Training Details

• Data Split: 8:1:1 (Train: 7,119 / Validation: 890 / Test: 890)
• Tokenizer: SentencePiece (KcELECTRA tokenizer)
• Max Length: 256
• Learning Rate: 3e-5
• Batch Size: 16
• Epochs: 6
• Early Stopping: Patience = 2
• Optimizer: AdamW
• Threshold Optimization: Independent per-label tuning (criteria = Micro-F1, Macro-F1)
  • Thresholds: [0.25, 0.675, 0.8, 0.75, 0.7]

---

Result

Metric	Value
Micro-F1	0.769
Macro-F1	0.758
Subset Accuracy	0.516

F1-score:

Emotion	F1-score
Dislike	0.72
Sympathy	0.81
Sadness	0.64
Surprised	0.80
Angry	0.82

---