BDRC/tibetan-ocr-benchmark

BDRC Tibetan OCR Benchmark (open subset)

A hand-transcribed benchmark for evaluating Tibetan OCR across writing styles and technologies. This is an open-access set of 472 page images with ground-truth transcriptions and per-page metadata (script, technology, legibility).

Companion to the model BDRC/tibetan-ocr and the leaderboard (dozens of OCR systems scored on this benchmark).

The transcriptions were produced by Dharmaduta.

The images were selected through detailed research on Tibetan writing styles in collaboration with Pentsok W. Rtsang who selected and cataloged the images. Most writing styles are represented in the benchmark, except those only present in images that cannot be made fully open access.

Contents & schema

One row per page. We deliberately keep only fully-documentable metadata; the internal fine-grained numeric script id and the 8-class script label are not published here.

column	description
image	the page scan (embedded bytes; anonymized filename)
transcription	Tibetan Unicode ground truth (main text box only), botok-normalized
id	stable per-page id (the anonymized page number, e.g. 0002)
technology	how the page was produced: manuscripts / blockprints / modern / metal_type / typewriter / …
script	coarse writing system: Uchen / Ume
script_4	writing-style family: Uchen / Pedri / Gyuyig / Druma / Tsugdri / Danyig / Non-applicable
period	palaeographic period of the writing style (e.g. Medieval (15)-present)
rarity	how common the writing style is (e.g. 4 - very common)
legibility	annotator legibility grade of the page (e.g. 2 - Fair, 4 - Excellent)

from datasets import load_dataset
ds = load_dataset("BDRC/tibetan-ocr-benchmark", split="test")

Transcription conventions

• Only the main text box is transcribed (no marginalia or pagination).
• Orthographic shorthands are kept as written.
• Placeholders in ground truth: I illegible · B broken paper · S struck-through · D/K dakini script · O ornamental. Scribal inserts are marked with ().

See transcription_conventions.md and analysis_conventions.md in this repo for the full manuals.

How CER is computed

Character Error Rate is computed after normalization, both sides:

1. Normalize with botok (hypothesis; ground truth is pre-normalized).
2. Strip all whitespace.
3. Fold repeated tsheg (U+0F0B) to one.
4. Remove placeholder characters (K O B I S).
5. CER = Levenshtein(hyp, ref) / len(ref) (rapidfuzz).

Aggregate CER is an unweighted mean over pages.

License

• CC0-1.0 for the transcriptions and metadata. Image-rights: BDRC claims no copyright over the scans; verifying the underlying works' copyright is the user's responsibility.

Citation

@misc{bdrc_tibetan_ocr_benchmark_2026,
  title  = {BDRC Tibetan OCR Benchmark},
  author = {Buddhist Digital Resource Center},
  year   = {2026},
  howpublished = {Hugging Face},
  note   = {https://huggingface.co/datasets/BDRC/tibetan-ocr-benchmark}
}

Funded by the Khyentse Foundation ("The BDRC Etext Corpus").