BDRC/tibetan-quotation-detection

Tibetan Quotation Detection Benchmark

254 classical Tibetan books with 23,095 human-reviewed quotation spans, split by book into train / validation / test.

Character offsets are inclusive on both ends: the span is text[start:end+1].

Source

Classical Buddhist commentaries digitized with support from the Tsadra Foundation and OpenPecha, annotated in two batches: an old batch (125 books, Quotation layer) and a new batch (195 books, Citation layer). Both layers mark the same thing — one text quoting a scripture, a named teacher, or a stock phrase — so they are merged into one label.

Cleaning

1. Human review. 398 books reviewed; 78 rejected whole for bad boundaries; 254 boundaries repaired in accepted books.
2. Span merging. The new batch often splits one quotation across lines. Pieces separated only by 0–2 spaces, newlines, or a shad (and not already closed) were joined: 28,892 → 23,476 spans.
3. Dropping under-annotated books. Quote markers appear at the same rate in every book (~1,200 per million characters), but some books had almost none of them annotated. Books with quotation density below 1% were dropped: 66 books, 22.5% of characters, only 1.6% of spans.

Quote types

Each span gets a quote type from the words that introduce it:

Type	Spans	Share	Introduced by
TEXT	13,164	57.0%	ablative markers (ལས, ནས) naming a text
MISC	4,354	18.9%	no marker found, or marker separated from the span
PERSON	3,636	15.7%	ergative markers or speech verbs (ན་རེ, གསུངས) naming a speaker
FORMULA	1,007	4.4%	set phrases (སྐད་དུ, ཇི་སྐད་དུ)
CHAINED	934	4.0%	follows a previous closer (ཞེས་དང, ཅེས་དང)

90.6% of spans end with an explicit closing marker.

Splits

Splits are book-level: a book's spans never appear in more than one split.

Books are stratified by quote type (TEXT / PERSON / MISC), so each split has the same mix of quote types. Density band, source batch, and overall quotation density are matched as well. FORMULA and CHAINED are too small to stratify on and come out slightly above target.

Split	Books	Spans	TEXT	PERSON	MISC	Density	Old batch
train	210	18,753	57.0%	15.8%	18.9%	7.40%	40.0%
validation	25	2,005	57.0%	15.7%	18.7%	7.37%	39.3%
test	19	2,337	57.1%	15.4%	18.8%	7.50%	38.9%

Leakage

• No text is shared between splits (the split is by book).
• The same quotation can appear in different books. Commentaries quote the same scriptures, so 724 of 4,342 held-out spans (16.7%) have text that also appears in train. These are kept, not removed, and every span carries a seen_in_train flag.

Report two scores: overall, and seen_in_train == false only. The gap shows how much the model relies on memorised quotes.

Known limits

• Some small boundary errors remain in accepted books.
• Some real quotations are still unlabeled, mostly in books with 1–3% density.
• Density (7.4%) reflects well-annotated books, not typical Tibetan text.

Loading

from datasets import load_dataset
ds = load_dataset("karma689/tibetan-quotation-detection")

Windowed configs (for mmBERT training)

Pre-tokenized 8192-token windows built with jhu-clsp/mmBERT-base. This is derived data — rebuild it if the tokenizer, window length, or step changes, and never mix windows from different builds.

Config	Labels
windowed_w8192_s4579	O=0, B-QUOTE=1, I-QUOTE=2
windowed_w8192_s4579_bioe	same, plus E-QUOTE=3 on the last token of each span

All quote types collapse to one QUOTE label. Load the parquet files directly (plain load_dataset(repo, config) picks the JSON builder from the default config):

from datasets import load_dataset
cfg = "windowed_w8192_s4579"   # or windowed_w8192_s4579_bioe
base = f"hf://datasets/karma689/tibetan-quotation-detection/{cfg}"
ds = load_dataset("parquet", data_files={
    s: f"{base}/{s}/*.parquet" for s in ["train", "validation", "test"]
})

The BIO config is also pinned at tag w1.0.

Geometry. Window 8192 tokens, step 4579, overlap 3613 (44%). In the name, s4579 is the step; the HuggingFace tokenizer argument is stride=3613. The overlap equals the longest span (3613 tokens).

Ownership. Each overlapped token is labeled in only one window (the first); elsewhere it is -100. Joining each book's owned tokens in window_index order rebuilds the book exactly, so stitching for inference or scoring is safe.

BIOE rule. An I token becomes E when the next owned token is not I, checked across window seams. Mapping E → I gives back the BIO config exactly. The shortest span is 3 tokens, so no single-token tag is needed.

Split	Windows	Spans
train	6,195	18,753
validation	645	2,005
test	657	2,337

• 513 spans cross a window seam.
• Token balance (owned tokens): O 92.37%, I 7.50%, B 0.07%, E 0.07% (in BIO, I is 7.57%).
• Tokenizer sha256 (combined): 9818cf906b79a54b964992d2f76db685ae8b9b3824e28a8f4e4f16425d026f9d

Citation

@misc{tibetan_quotation_benchmark,
  title  = {Tibetan Quotation Detection Benchmark Dataset},
  author = {karma689},
  year   = {2026},
  url    = {https://huggingface.co/datasets/karma689/tibetan-quotation-detection}
}

Acknowledgements

Source texts were digitized and made available by the Buddhist Digital Resource Center (BDRC). We gratefully acknowledge BDRC. Annotations were prepared through OpenPecha with support from the Tsadra Foundation.