HTmiR — Kraken 7 HTR (CATMuS Medieval French, 13th c.)
This repository hosts a fine-tuned Kraken 7 recognition model for Medieval French (13th century) handwriting.
The model file is intended to be used with the Kraken CLI:
kraken ... ocr -m htmir-french-13c_best.safetensors
Files
htmir-french-13c_best.safetensors — Kraken 7 recognition weights
Training / validation metrics
Best validation (early-stopped):
- Char accuracy:
0.95533
- CER:
0.04467
- WER:
0.236
- Word accuracy:
0.764
Dataset split sizes (CATMuS/medieval):
- Train:
19238
- Validation:
969
- Test:
1369
Usage with Kraken 7
Line OCR (pre-segmented lines)
kraken -i line.png out.txt ocr -m htmir-french-13c_best.safetensors -sPage OCR (segmentation + OCR)
kraken -i page.png out.xml -x segment -bl ocr -m htmir-french-13c_best.safetensorsNotes
The training/evaluation pipeline recorded metrics from validation logs (Kraken 7 binary test evaluation had known issues on this setup).
