Kiuyha/indobert-job-fraud-detection

IndoBERT Job Fraud Classifier (v1)

Model Description

Fine-tuned indobenchmark/indobert-base-p2 for detecting fraudulent job postings in Indonesian language text.

Training Configuration

• —Max Sequence Length: 512
• —Learning Rate: 2e-5 (AdamW, weight decay 0.01, warmup ratio 10%)
• —Batch Size: 16
• —Max Epochs: 4 (Early stopping patience = 1 on validation macro-F1)
• —Seed: 42
• —Class Balancing: Loss weight ratio + minority class duplication per batch

Evaluation Metrics (Hold-out Test Set)

• —Accuracy: 0.9940 (Target >= 0.85)
• —Precision (Fraud): 0.9684 (Target >= 0.88)
• —Recall (Fraud): 1.0000 (Target >= 0.82)
• —Macro F1: 0.9902 (Target >= 0.85)

Deployment Status

model_artifacts.deployed=false (activate manually, then restart the api service)