AvoCahDoe/morocco-bo-ocr-latin-v3-finetuned

Morocco BO OCR — Latin PP-OCRv3 Fine-tuned

Fine-tuned latin_PP-OCRv3_mobile_rec on Moroccan Bulletin Officiel line crops (dataset). Better suited for French/Latin BO text.

Training summary

Metric	Value
Best val accuracy	0.9605
Peak eval val acc (log)	0.9605
Early stopped	True
Best eval index	68
Base model	latin_PP-OCRv3_mobile_rec
Profile	latin

Usage (PaddleOCR)

from paddleocr import TextRecognition

rec = TextRecognition(
    model_dir="path/to/inference",
    model_name="latin_PP-OCRv3_mobile_rec",
)
result = rec.predict("line_crop.png")
print(result)

Repository layout

inference/          # Exported Paddle inference model
checkpoints/        # best_accuracy training checkpoint
metrics/            # training + eval JSON metrics
plots/              # training and evaluation plots
config.yml          # PaddleOCR training config snapshot

Training artifacts

See metrics/ and plots/ for full learning curves, early-stop state, and test-set OCR metrics.