DC-D5-LegalBERT v1: Deference Classifier — Structural Rationale for Restraint
Fine-tuned from nlpaueb/legal-bert-base-uncased for 4-class classification of ECtHR paragraphs
on deference dimension D5 (Structural Rationale for Restraint).
Classes: ND (0) | MD (1) | HD (2) | NA (3)
Scoring rule: continuous deference score = 0×p(ND) + 0.5×p(MD) + 1.0×p(HD);
NaN if p(NA) is argmax.
Training data snapshot (v1)
- Train paragraphs : 549
- Val paragraphs : 164
Validation performance (v1)
- Best macro-F1 : 0.8239
- Mean macro-F1 : 0.7496 ± 0.0456
- All seeds : [0.8239, 0.6903, 0.7574, 0.7609, 0.7155]
Training setup
- Regime : small (10 epochs, batch size 8, 5 seeds)
- Learning rate : 2e-05
- Seeds tested : [0, 1, 2, 3, 4]
Intended use
Part of a Master's Thesis measuring judicial deference in ECtHR Article 8 migration cases.
Not intended for production use.
