todobichig/tddata

todo-bichig-data — line-level OCR corpus for Todo Bichig (Clear Script)

17,604 line images (vertical column crops) from Kalmyk/Oirat manuscripts and scholarly facsimile editions, each paired with a Todo Bichig Unicode transcription. This is the corpus behind the Zaya Pandita OCR system and, to our knowledge, the first source-aware line-level recognition benchmark for the script.

Layout

final_lines_todo.txt   manifest: <text>\t<image id>   (17,604 entries)
lines.tar              line crops (17,604 PNG/JPG, vertical orientation) — unpacks to lines/
splits/                canonical benchmark manifests (see below)

• Transcriptions are Todo Bichig Unicode (Mongolian block); @ marks a page/section boundary (present in 115 lines).
• Image ids in the manifest resolve to files in lines/ after extraction.
• The crops ship as one uncompressed tar because Hugging Face limits a repository directory to 10,000 files.

Canonical splits (splits/)

File	Lines	Purpose
train.txt	14,909	training
val.txt	1,713	validation / model selection
test.txt	671	held-out benchmark — never used for tuning
train_25pct/50pct/75pct.txt	—	data-scaling curve subsets
lomo/	—	leave-one-manuscript-out splits
test_legacy.txt	671	pre-repair test references (historical comparison only)

Protocol highlights:

• Source-aware: fragments of one source never cross the train/test boundary; test lines are excluded from training pools by filename and by normalized text.
• Page-disjoint train/val: validation is assembled from whole pages; scan-duplicate pages are linked before splitting.
• The canonical splits cover 17,293 lines: 311 cross-edition duplicates of test lines are excluded from every split.
• Scoring: training.ocr_common.score_predictions in the code repository (macro CER/WER/exact-match, plus normalized variants: x≡ᡍ, @≡space).

Best published baseline: TrOCR-Base + rot90 + augmentation, CER 2.15 ± 0.30 % on test.txt (todobichig/tbocr).

Composition

• 8,066 lines from the initial transcribed collection (starting point of the project, provided by Badma Ontaev);
• 9,538 lines mined from five scholarly facsimile editions with parallel printed transliterations: pages were rendered, column crops aligned to the printed text, and candidates admitted at a similarity threshold of sim ≥ 0.8 (16,690 candidates → 9,538 admitted).

Use with the Zaya Pandita OCR code

hf download todobichig/tddata --repo-type dataset --local-dir data/final_data
tar -xf data/final_data/lines.tar -C data/final_data && rm data/final_data/lines.tar

This reproduces exactly the layout the training code expects (data/final_data/lines/ + final_lines_todo.txt); the canonical splits are also shipped in the code repository under data/splits/. Training/evaluation protocol: github.com/nfrvnikita/ZayaPanditaOCR, training/README.md.

Access

Private dataset for the Zaya Pandita OCR project (HSE University / Kalmyk State University collaboration). Contact the authors before any redistribution.