BDRC/tibetan-ocr-diagnostic-benchmark

Tibetan OCR Diagnostic Benchmark (OFAT)

A small, controlled diagnostic OCR benchmark of 300 synthetic Tibetan pecha-page images with exact, noise-free ground truth. It is a scientific instrument for measuring how OCR character error rate (CER) responds to individual difficulty factors one at a time (OFAT) — not a coverage-maximizing training set.

Code & regeneration: https://github.com/buda-base/synthetic-ocr-benchmark-tools (diagnostic_benchmark/)

What's inside

Each row is one rendered page:

column	meaning
image	grayscale JPEG page (viewable in the dataset viewer)
transcription	exact LuaLaTeX-rendered ground truth (line breaks as \\n), Unicode-normalized for bec-ocr-training scoring
file_name	stable id, e.g. axis_length_bin0_0161.jpg
script	uchen
technology	synthetic
benchmark_group / axis	difficulty axis (stack_rarity, repetition, length, font_size)
axis_bin	ordinal bin (0 = easiest / center for that axis)
axis_value	numeric level (rarity score, target rep_score, syllables/page, or font size in pt)

Also in the repo (not in the Arrow split): manifest.csv (per-image provenance) and provenance.json (full build config).

Fixed center configuration

All pages share one Uchen font (Jomolhari-id), one page width, one image width (2000 px), no augmentation beyond minimal raster naturalization, and a coverage gate from coverage_report/out/stack_support.parquet so every stack is known to render faithfully.

Axes (OFAT sweeps)

• stack_rarity — common → rare Tibetan stacks (5 bins × 16)
• repetition — char-20-gram rep_score from ~0 → 0.6 (5 × 16)
• length — syllables/page from 15 → 1206 (longest BEC train page + 20%; 5 × 16)
• font_size — 12 / 18 / 28 pt at fixed pixel width (3 × 20)

Load

from datasets import load_dataset

ds = load_dataset("BDRC/tibetan-ocr-diagnostic-benchmark", split="test")
print(ds[0]["file_name"], ds[0]["axis"], ds[0]["axis_bin"])
ds[0]["image"]  # PIL.Image
print(ds[0]["transcription"][:200])

Filter one axis:

length = ds.filter(lambda r: r["axis"] == "length")

Source & rights

• Text: OpenPecha BoCorpus (public domain), rendered with LuaLaTeX / HarfBuzz
• Images: synthetic, released under CC0 1.0 (public domain dedication)
• Please cite BDRC, the BDRC Etext Corpus / OpenPecha BoCorpus, and OpenPecha Tibetan Fonts
• Builder: https://github.com/buda-base/synthetic-ocr-benchmark-tools/tree/main/diagnostic_benchmark

Citation

@misc{bdrc_tibetan_ocr_diagnostic_benchmark,
  title  = {Tibetan OCR Diagnostic Benchmark (OFAT)},
  author = {Buddhist Digital Resource Center (BDRC)},
  year   = {2026},
  url    = {https://huggingface.co/datasets/BDRC/tibetan-ocr-diagnostic-benchmark},
  note   = {Code: https://github.com/buda-base/synthetic-ocr-benchmark-tools}
}