johnlockejrr/ppocrv6-sam-heb

PP-OCRv6 Samaritan Hebrew recognizers

Line-level text recognition models for Samaritan Hebrew manuscripts, trained with kraken’s PP-OCRv6 port (ketos train --arch ppocrv6) and usable with oxygraphos-ocr / Ppocrv6-OCR (ppocrv6-ocr) or stock kraken ≥ 7.0.

These are recognizers only (line crops / baseline polygons). Pair with a page segmenter for full-page OCR.

Files

File	Variant	Val char accuracy*	Codec	Size (approx.)	Notes
ppocrv6_sam_heb_small.safetensors	small	0.9872	28 Hebrew letters + space	~12 MB	From scratch
ppocrv6_sam_heb_medium.safetensors	medium	0.9877	28 Hebrew letters + space	~60 MB	From scratch
ppocrv6_sam_heb_medium_ft.safetensors	medium	0.9886	~1622 chars (multilingual head retained)	~61 MB	Fine-tuned from a multilingual PP-OCRv6 checkpoint

*Accuracy is the kraken checkpoint score embedded in the original best_*.safetensors filename (validation character accuracy).

Default recommendation: ppocrv6_sam_heb_medium_ft.safetensors for best accuracy on this domain. Use small when latency / footprint matters. Prefer the Hebrew-only small / medium codecs if you want a locked Samaritan alphabet (no Latin / punctuation emissions).

Model details

Architecture	PPOCRv6Model (kraken), task recognition
Input	RGB line images, target height 96, baseline / polygon-dewarped lines (seg_type: baselines)
Normalization	NFD + whitespace normalize; BiDi reordering enabled at train time
Data	Samaritan Hebrew PAGE-XML line corpus (sam_44_mss_pango / additional split), compiled to binary Arrow
Min kraken	7.0.0 (_kraken_min_version in metadata)

Codec is stored inside each .safetensors file under kraken_meta — no sidecar alphabet file is required.

Usage

Download

hf download johnlockejrr/ppocrv6-sam-heb ppocrv6_sam_heb_medium_ft.safetensors --local-dir ./models

oxygraphos-ocr / ppocrv6-ocr

ppocrv6-ocr -d cuda:0 infer \
  -m ./models/ppocrv6_sam_heb_medium_ft.safetensors \
  -f page --output-dir ./out page.xml

kraken

kraken -i line.png text.txt ocr -m ./models/ppocrv6_sam_heb_medium_ft.safetensors

Intended use

• Samaritan Hebrew manuscript line recognition (square Hebrew Unicode)
• Offline HTR pipelines that already produce line polygons / crops

Limitations

• Not a page layout / line detector
• Trained on this manuscript family; other hands / scripts may need fine-tuning
• medium_ft retains a large multilingual output vocabulary; greedy decoding can occasionally emit non-Hebrew symbols on ambiguous lines
• Hebrew-only small / medium cannot emit characters outside their 28-letter (+ space) codec

Training summary

• Trainer: kraken ketos / PP-OCRv6 (AdamW+Muon, cosine schedule, bf16-mixed)
• Augmentation on; --base-dir R / RTL-aware setup for Hebrew
• medium / small: trained from scratch on the Samaritan binary dataset
• medium_ft: fine-tuned from a multilingual medium checkpoint onto the same data (lower LR); final weights keep the large pretrained codec

Citation / credits

• PP-OCR / PP-OCRv6: PaddlePaddle/PaddleOCR
• kraken PP-OCRv6 port: mittagessen/kraken
• Packaging: johnlockejrr/oxygraphos-ocr (Ppocrv6-OCR)

License

Apache-2.0 (weights and card). Upstream PaddleOCR / kraken licenses still apply to those codebases.