albertklorer/single-page-image-markdown-1m

Single-Page Image-Markdown 1M

Cleanup (2026-10-21)

The original release contained exactly 1,000,000 rows. After an exporter defect was fixed, 22,670 arXiv rows were removed: 17,201 rows with exact stored signatures of injected Markdown blockquote/code-fence markers, plus 5,474 rows matching exact source SVG picture titles that had been merged into body text (5 rows overlapped both classes). The current dataset therefore has 977,330 rows: 820,513 from the arXiv source and 156,817 from SafeDocs. A full post-cleanup census found zero remaining removal IDs, zero exact injected-marker blocks, and zero matched picture-title rows under those measured signatures.

977,330 single-page training rows — one page image plus that page's Markdown — formed by unifying and shuffling two verified datasets:

source	rows
albertklorer/arxiv-varied-single-pages-en80-zh20	820,513
albertklorer/safedocs-perfect-single-pages-markdown	156,817

Language totals: 820,135 English (en) and 157,195 Chinese (zh) per the strict per-page majority labels of the sources. Rows were originally an exact concatenation of the two sources, deterministically shuffled with seed string single-page-image-markdown-1m/shuffle-v1 (BLAKE2b rank over source/shard/row/id) and re-sharded into 1,000 parquet shards (now at most 1,000 rows each after cleanup) under data/train/.

One documented deviation from byte-identical copies: 15 SafeDocs rows whose blocks objects carried a ## prefix in their text field (their markdown column was already single-#) had that prefix flattened to a single #, so headings are single-# in both the markdown and blocks columns throughout.

Schema

Identical columns in every row:

column	type	meaning
id	string	source row id (unique across the union)
image	struct{bytes, path}	rendered/scanned page image
markdown	string	the page's Markdown (single-# headings; no HTML except tables; figures as , 0–1000 coords)
language	string	en or zh
source_dataset	string	provenance repo (arXiv set, or a SafeDocs judged source)
zip	string	source document id (arXiv id for arXiv rows)
pdf	int32	variant/index within source packaging
page	int32	page index within its rendered document variant
page_count	int32	pages in that variant
width, height	int32	image dimensions
judge_verdict, judge_reason	string	upstream verification fields
blocks	string	JSON blocks for the page (see note)
paddleocr_result	string	upstream OCR field (may be null)

Note on blocks: arXiv rows store a JSON list of Markdown strings; SafeDocs rows store a JSON list of objects ({bbox, label, order, text}) with 0–1000 bounding boxes. The source_dataset column tells them apart.

Verification

A full census over all 1,000 shards confirmed: 1,000,000 rows; every shard exactly 1,000 rows; en = 829,667, zh = 170,333; source counts 843,183 + 156,817; the (source, id) multiset hash is identical to the pre-shuffle census of the two sources (no row lost or duplicated); uniform schema; zero multi-# headings in markdown (outside code fences) or blocks; and, over a 3,000-row sample, zero language-label mismatches, zero bad figure references, and zero bad images.

Source datasets

• albertklorer/arxiv-varied-single-pages-en80-zh20 — template-varied renders of native-HTML arXiv documents; pages exported only with 100% printed-character coverage and no cross-page dependencies.
• albertklorer/safedocs-perfect-single-pages-markdown — perfect single pages (image + Markdown) judged from SafeDocs.

Content rights remain with the original documents' authors/licensors.