Agreemind/banking-bert-turkish

Turkish BERT fine-tuned for multi-label risk classification in Turkish consumer banking contracts. Detects 14 risk categories including hidden fees, collateral clauses, default escalation, and more.

Performance

Evaluated via 5-fold document-level cross-validation on 80 Turkish banking contracts (7,020 clauses).

Metric	Score
Micro-F1 (CV)	0.6657
Macro-F1 (CV)	0.6226

Per-class F1 (5-fold CV)

Category	F1
hidden_fees	0.76
dispute_limitation	0.61
broad_collateral	0.74
default_escalation	0.71
account_freeze	0.65
currency_risk	0.68
unilateral_terms_change	0.57
unilateral_rate_change	0.72
data_sharing	0.69
auto_enrollment	0.42
cross_default	0.44
overdraft_penalty	0.51
early_payment_penalty	0.64
bundled_insurance	0.59

Training

• Base model: dbmdz/bert-base-turkish-cased
• Loss: Multi-label Focal Loss (gamma=2.0, alpha=0.75)
• Optimizer: AdamW, lr=2e-5, weight_decay=0.01
• Epochs: 10
• Training data: 80 Turkish banking contract PDFs, 7,020 clauses (Akbank, Garanti, İşbank, DenizBank, Halkbank, VakıfBank, QNB, TEB, YapıKredi, Ziraat, ING, KuveytTürk, Şekerbank, HSBC Turkey, Odea Bank, Albaraka Türk, Türkiye Finans, Burgan Bank and more)
• Threshold: Fixed 0.5 per label

Usage

License

MIT