Key-value pair extraction on OCR-derived clinical reports (test)
84.9F1 (EM)BERT-based QA model
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| BERT-based QA modelTop-N=Top95, Key Coverage=90.07%2026.05 | 84.9 | 89.9 | 92.3 | 97.8 | 79.6 | 84.4 | |
| BERT-based QA modelTop-N=Top100, Key Coverage=90.64%2026.05 | 84.1 | 89.6 | 91.2 | 97.2 | 79 | 84.2 | |
| BERT-based QA modelTop-N=Top90, Key Coverage=89.43%2026.05 | 83.9 | 89.3 | 91.5 | 97.4 | 78.5 | 83.5 | |
| BERT-based QA modelTop-N=Top80, Key Coverage=87.89%2026.05 | 83 | 88.3 | 91.4 | 97.3 | 77.2 | 82.1 | |
| Qwen3-0.6B-SFTKey Coverage=100%2026.05 | 81.4 | 95.7 | 85 | 96.5 | 78.1 | 94.9 | |
| BERT-based QA modelTop-N=Top50, Key Coverage=77.85%2026.05 | 78.6 | 83.6 | 91.2 | 96.9 | 70.5 | 75.1 | |
| BERT-based QA modelTop-N=Top20, Key Coverage=53.62%2026.05 | 62.7 | 66.8 | 90.2 | 96.2 | 49.3 | 52.6 | |
| BERT-based QA modelTop-N=Top10, Key Coverage=38.65%2026.05 | 51.3 | 53.9 | 91.7 | 96.8 | 36.8 | 38.8 |