formatting-title-author-parquet (private)

Raw-text version of the title/author data: 37,101 works, each with one title span and one author span. For review and sharing; the tokenized training sets are separate.

What a row is

One work (a text segment inside a BDRC volume). The text is the first and last ~3,000 characters of the work, each edge moved to a line break; a work of 6,000 characters or fewer is kept whole. A span is never cut.

column	meaning
work_id	file name + segment (v3: prefix for BDRC v3 works)
text_head, text_tail	first / last ~3,000 characters (text_tail is empty when whole_work is true)
work_chars	length of the full work
title, author	span text
title_region, author_region	head or tail: which text the offsets refer to
title_start/end, author_start/end	character offsets inside that text (end exclusive)
source	tei, google_books, google_vision, paddleocr_v1, paddleocr_v2, ocr_qwen_v1, ocrv1-ws-ldv1, v3_auto
book_type, bdrc_work_id	from the BDRC catalog (book_type is unknown when BDRC has none)

Counts

source	train	validation	test	total
v3_auto	7,812	0	0	7,812
tei	6,164	539	486	7,189
paddleocr_v2	5,117	637	640	6,394
google_books	3,829	379	350	4,558
ocr_qwen_v1	3,208	309	342	3,859
google_vision	3,008	304	225	3,537
ocrv1-ws-ldv1	2,012	207	226	2,445
paddleocr_v1	1,063	124	120	1,307
all	32,213	2,499	2,389	37,101

book_type	train	validation	test	total
Manuscript	13,683	860	835	15,378
Modern	11,710	1,154	1,042	13,906
Relief_WoodBlock	6,494	465	497	7,456
unknown	326	20	15	361
all	32,213	2,499	2,389	37,101

Sources (source column)

source says where the text of the work came from. The title/author spans were marked by human annotators in every row except v3_auto, whatever the text source. Rows with source = v3_auto have automatic spans (found by string matching, ~91-94% right); all other rows have annotator spans.

source	text origin
tei	BDRC e-text (TEI format), human-typed digital text
google_books	Google Books text source
google_vision	Google Vision text source
paddleocr_v1, paddleocr_v2	OCR with the PaddleOCR engine (two versions)
ocr_qwen_v1	OCR text (ocr_qwen_v1 pipeline)
ocrv1-ws-ldv1	OCR text (ocrv1-ws-ldv1 pipeline)
v3_auto	PaddleOCR text from the BDRC char-level outlines release v3; spans found automatically

How the split was made

• The pool is all 34,774 annotated works. Copies of the same work (same typed title+author, same first 300 characters, or a near-identical text start) were reduced to one: 30,009 unique annotated works, of which 29,297 appear here (the rest have no author span left after the title-wins rule or a span lost in the cut middle).
• Where copies were found, the higher-ranked source was kept: tei, google_books, google_vision, paddleocr_v2, ocrv1-ws-ldv1, ocr_qwen_v1, paddleocr_v1; then the longer text.
• Split into train / validation / test about 80/10/10, stratified on source x book type x length and spread over collection families. Duplicate works never sit in different splits; a collection can have works in more than one split (447 of 2,896 do).
• Validation and test are annotator spans only. The 7,812 BDRC v3 works (source = v3_auto, PaddleOCR text from the BDRC char-level outlines release v3) are in train only.

Known limits

• One title and one author per work; other mentions and second authors are not marked.
• Span edges vary between annotators (e.g. whether the closing shad or བཞུགས་སོ is included).
• Duplicate detection looks at the first 1,500 characters; some copies can slip through, and some different works can be wrongly merged.