BDRC/tibetan-spelling-correction-dataset

Tibetan Spelling Correction

Sentence-level spelling correction pairs for Tibetan. Each row pairs an annotator's transcription of a manuscript page segment with the final reviewer's corrected version, from the double-annotation workflow of the BDRC Etext Corpus.

For training and evaluating post-correction models such as TiSpell.

Source batches: Ume 1-4, Uchen 1-4. 4,672 pages.

Contents

	train	validation	test	All
error pairs	43,015	2,268	2,229	47,512
clean pairs	47,370	2,564	2,584	52,518
total	90,385	4,832	4,813	100,030

Clean pairs, where annotator and reviewer agree, are included so a model learns not to alter text that is already correct.

Error categories

category	count	share
syllable_level	23,219	48.9%
single_char_substitution	11,067	23.3%
large_rewrite	6,474	13.6%
single_char_insert_delete	4,205	8.9%
punctuation_only	2,540	5.3%
multi_edit	7	0.01%

punctuation_only covers tsheg differences within the text, where a missing tsheg between syllables is an orthographic error. Differences that were only a terminal shad were treated as formatting and removed.

Script

Ume accounts for roughly two thirds of error pairs while the two scripts are close to balanced among clean pairs, so Ume produces more transcription errors per page. Report per-script results rather than a blended figure.

Splits

Page-stratified. Every page appears in exactly one split, so no segment is shared and both annotators' versions of a page stay together.

split	pages
train	4,182
validation	232
test	230

No exact source overlap between splits. Near-duplicates on error pairs: 0.17% at edit distance 1, 2.97% at distance 5.

Schema

column	type	description
page_id	string	{batch}/{image}, the split unit
source	string	annotator transcription
target	string	final reviewer transcription
is_error	bool	true when source and target differ
diff_category	string	edit type; empty for clean pairs
edit_distance	int	Levenshtein; 0 for clean pairs
high_confidence_error	bool	both agreed and the reviewer still corrected it
batch_name	string	source batch

Preprocessing

1. Normalize: Unicode NFC, whitespace, tsheg and shad spacing. Before comparison, so encoding differences are not counted as errors.
2. Segment the reviewer transcript on shad.
3. Transfer those boundaries onto the annotator text with fast_antx.
4. Keep segments where the reviewer side ends in a shad. Minimum 8 characters.
5. Drop pairs below 0.55 similarity, these are misaligned spans.
6. Remove transcription artifacts: Latin placeholders, ornament marks, leading segmentation debris, editorial brackets, non-Tibetan punctuation.
7. Deduplicate; drop pairs over 380 characters.
8. Drop pairs whose only difference is a terminal shad, as formatting rather than spelling.
9. Split 90/5/5 by page.

Sequence lengths: median 29 tokens, max 363, under the 512-token limit of openpecha/tibetan_RoBERTa_S_e3.

Load

from datasets import load_dataset

ds = load_dataset("BDRC/tibetan-spelling-correction-dataset")
errors = ds["train"].filter(lambda r: r["is_error"])
gold = ds["test"].filter(lambda r: r["high_confidence_error"])

The clean-pair ratio to use in training is left to the user; the full distribution is published.

Limitations

The final reviewer's transcription is taken as correct.

Citation

@misc{bdrctibetanspellcorrection,
  title  = {Tibetan Spelling Correction Dataset},
  author = {Buddhist Digital Resource Center and OpenPecha},
  year   = {2026},
  url    = {https://huggingface.co/datasets/BDRC/tibetan-spelling-correction-dataset},
  note   = {Transcriptions from BDRC}
}

License

Transcriptions derive from the open access collection of the Buddhist Digital Resource Center. Not all source material is in the public domain; some is from recent publications possibly under copyright. We provide the transcriptions under the Fair Use copyright exception, and any reuse will have to be based on a copyright analysis. The alignment and annotation data is provided under CC0 1.0 Universal.

Acknowledgements

Transcriptions provided by the Buddhist Digital Resource Center (BDRC). Developed by Dharmaduta from specifications provided by BDRC for the project "The BDRC Etext Corpus", with funding from the Khyentse Foundation. Developed by Dharmaduta / OpenPecha.