vekamin/mrc-legalbert-v2

MRC-LegalBERT v2: Migration Relevance Classifier

Fine-tuned from nlpaueb/legal-bert-base-uncased for binary classification of ECtHR paragraphs as in-scope (Article 8 migration case) or out-of-scope.

Training data snapshot (v2)

• —IN cases: 72 (annotated in INCEpTION, not on exclusion list)
• —OUT cases: 106 (exclusion list)
• —Train split: 125 cases / 1626 paragraphs
• —Val split: 32 cases / 417 paragraphs

Validation performance (v2)

• —Macro F1: 0.9298
• —Recall-IN: 0.3939 (expected to improve with more annotation data)
• —Precision-IN: 0.8667

Training setup

• —Regime: large (5 epochs, batch size 32, 3 seeds)
• —Learning rate: 2e-05
• —Seeds tested: [0, 1, 2]
• —All F1s: [0.905, 0.9243, 0.9298]
• —Mean F1: 0.9197 ± 0.0106

Intended use

Part of a Master's Thesis measuring judicial deference language in ECtHR Article 8 migration cases. Not intended for production use.