Medical Prescription OCR TrOCR Small
This model was fine-tuned in Kaggle on line-level medical prescription crops
exported through data/splits/image_annotations.csv.
Base model
microsoft/trocr-small-handwritten
Dataset summary
- Splits root:
/kaggle/working/downloads/splits_extracted/splits - Train lines:
20250 - Validation lines:
2526 - Test lines:
2517
Training setup
- Epochs:
6 - Effective batch size:
24 - Learning rate:
4e-05 - Weight decay:
0.01 - GPUs seen:
2 - Validation beams:
2 - Final eval beams:
4
Metrics
- Best validation CER:
0.0056 - Test line CER:
0.0184 - Test full-prescription CER:
0.0128 - Test line word accuracy:
0.9756 - Test full-prescription word accuracy:
0.9865