PDF / multi-column order
Transformer-XL paper, page 1
A real ACL paper page with title, abstract, two-column body flow, equations, and a semantic successor score of 1.0 on the tracked partial sidecar.
PDF / PP-Structure evidence
Attention Is All You Need, page 1
A structure-enhanced replay of the canonical Transformer paper. Native PDF text remains the editable layer while PP-Structure contributes local semantic boundaries; 56 of 59 elements are editable across five runtime streams.
Page 1 of arXiv:1706.03762, rendered at the benchmark resolution.
The tracked partial sidecar retains semantic successor accuracy 1.0 after structure fusion.
PDF / table structure
BYD annual report, financial page 136
A Chinese A-share financial report page used to test dense rules, tabular cells, local row-major structure, and translation-local table streams.
Source page 136 from the public annual-report PDF; the full third-party PDF is linked rather than bundled.
The exported DOM keeps table cells and vector rules addressable instead of flattening the page into one image.
PDF / PaddleOCR-VL evidence
PUMA annual report, page 5
A mixed image-and-text corporate report page replayed with real PaddleOCR-VL 1.6 output. The model evidence matches 24 elements and contributes local block streams without replacing native PDF text.
PUMA 2024 annual report, page 5, sampled at 144 DPI.
The page exposes 25 editable elements and seven runtime streams at visual similarity 0.9645.
Image / OCR anchors
JD homepage screenshot
A dense portal screenshot used as a first-class image input. The current conversion produces 141 editable OCR anchors and four local reading streams.
The raster source stays visible while OCR and layout evidence become editable semantic anchors.
The fidelity layer preserves the source visual while the underlying DOM provides transparent editable OCR anchors.
Website / browser print PDF
Hacker News print, page 1
A real browser-printed portal page with ranked entries, metadata, and repeated link rows. The conversion keeps 92 editable elements in 12 local streams and scores 1.0 on the tracked successor sidecar.
A tracked Playwright print capture of the public Hacker News homepage.
The browser-print visual remains exact while repeated title and metadata rows stay independently editable.
Image / relation-order benchmark
ROOR validation image 86263525
A photographed document page with OCR anchors and external structure evidence. It exposes 28 editable elements across nine runtime streams; four uncertain block transitions remain review-only.
Fixed ROOR validation sample 86263525; the image is the first-class source.
Semantic successor accuracy remains 1.0; ambiguous cross-block handoffs are visible as review evidence rather than forced runtime order.
Image / answer-separated hierarchy input
Comp-HRDoc page 1710.06349_4
A graphical multi-column benchmark page converted from its image and answer-free text anchors. The Gallery build does not read paragraph membership or successor labels.
The bundled OCR wrapper contains 80 answer-free text boxes and no evaluation labels.
The fidelity page is visually exact and exposes 80 editable text anchors across five local streams.
PDF / three-column magazine
Hello World Magazine #22, page 5
A true three-column magazine contents page with mixed images and dense caption/footer content. It is the strongest candidate-disagreement stressor in the current external benchmark set, and the page that drove the guarded magazine column-flow fixes (labelled successor accuracy 0.22 to 0.78).
Hello World issue 22 (Raspberry Pi Foundation, free download), sampled at 144 DPI.
The fidelity layer keeps the magazine visual while the column-flow reading order exposes editable anchors across the three index columns, feature boxes, and footer artifacts.
PDF / float-dense two-column paper
Segment Anything, page 5
A CVPR two-column paper page with a full-width Figure 4 caption and cross-column floats. The tracked semantic sidecar scores pair, successor, and relation accuracy 1.0 on this page.
Segment Anything (arXiv:2304.02643), sampled at 144 DPI.
The fidelity layer preserves the dense figure layout while caption, heading, and two-column body anchors stay editable.
PDF / equation-heavy two-column paper
Mamba, page 4
A math and algorithm heavy two-column paper where display equations and algorithm boxes interrupt paragraph flow. It has the lowest reading-order confidence in the current paper set.
Mamba (arXiv:2312.00752, CC BY 4.0), sampled at 144 DPI.
The fidelity layer keeps equations and algorithm boxes intact while text anchors remain editable for translation experiments.