PHI de-identification on Technetium-I (test)
0.984Micro-F1ProtactiniumBERT-100M
Evaluation Results
| Method | Links | |
|---|---|---|
| ProtactiniumBERT-100MBackbone=Transformer encoder (12 layers, 12 attention heads, hidden size 768), Parameters=~100M, Tokenizer=WordPiece (30k, uncased), Domain adaptation=continued pretraining on Technetium-I train split, Task Head=token classification with BIO tagging2026.01 | 0.984 | |
| ClinicalBERT2026.01 | 0.976 | |
| BioBERT2026.01 | 0.972 | |
| Rule-based2026.01 | 0.895 |