Project: BDH-ECHR (Post-Transformer Legal AI)
Trained and evaluated an NLP post-transformer model for legal multi-label violation classification using an ECHR legal corpus. Produced performance metrics to compare against baseline transformer models and parameter efficiency. Conducted interpretability analysis by mapping neurons to legal concepts. • Trained Dragon Hatchling (BDH) on legal corpus for article-violation classes • Achieved Micro-F1 0.7716 at 20.4M parameters • Compared results vs BERT-base and matched Longformer with fewer parameters • Mapped 218 interpretable neurons across legal concepts and audited monosemanticity