Table extraction (TE) is a key challenge in document understanding. Traditional approaches detect tables first, then recognize their structure. Recently, interest has surged in developing methods, such as vision-language models (VLMs), to extract tables directly in their full page or document context. However, a lack of annotated data has made progress difficult to demonstrate. To address this, we create a new large-scale dataset, PubTables-v2. PubTables-v2 unifies TE across various levels of surrounding context and, notably, is the first benchmark for multi-page TE. Our evaluations reveal that while current frontier models strongly outperform (+0.354GriTSCon) small models on the most complex task (full-document multi-page TE), this gap can be closed or even reversed (−0.056GriTSCon) on narrower tasks (cropped table extraction) with targeted training. Data is available at https://huggingface.co/datasets/kensho/PubTables-v2.
Figures & tables
Figure 1: Among the more than 500k tables in PubTables-v2 are 9,492 multi-page tables, by far the largest set to date. PubTables-v2 includes tables spanning up to 13 pages.
Figure 2: The PubTables-v2 Single Pages subset contains 548,414 tables in a full-page context. Above are four pages annotated with bounding boxes for table structure plus hierarchical relationships annotated with arrows between tables, their captions, and their footers.
Dataset Name
Total
Long
Wide
FinTabNet
113k
7,604
477
PubTables-1M
948k
72,570
31,987
PubTables-v2
136k
100,781
37,106
Table 1: Comparison of long and wide cropped tables in document TSR datasets.
Dataset
Year
Pages w/ Structure
Page-Level Tables
Full Docs w/ Structure
Multi-Page Tables
No Missing Tables
Cell Bbox
ICDAR-2013 [ 7 ]
2013
206
250
0 §
0
✓
✓
ICDAR-2019 cTDaR [ 4 ]
2019
248
335
0
0
✓
✓
FinTabNet [ 33 ]
2021
76,077
112,887
0
0
✗
✓
CTE [ 6 ]
2023
< 35,000
∼ 35,000
0
0
△ ‡
✓
OmniDocBench v1.5 [ 16 ]
2025
458
665
0
0
✗
✗
PubTables-1M [ 23 ]
2021
0 †
0 †
0
0
— †
— †
Table 2: Comparison of datasets for page-level and document-level table extraction. Page-Level Tables: tables annotated with structure in full-page context (not cropped). Pages w/ Structure: pages with ≥ 1 such table. No Missing Tables: whether all tables on annotated pages are guaranteed to be included (✓= guaranteed; ✗= known missing; △ = unlikely but not guaranteed). PubTables-v2 is the first dataset to provide page-level table structure annotations at scale with a completeness guarantee, and the first with multi-page tables.
Figure 3: PubTables-v2 contains 136k long and wide cropped tables for TSR in the same format as PubTables-1M (see Tab. 1 for a comparison). Depicted above are structure annotations with bounding boxes for a wide table with 21 columns.
Figure 4: PubTables-v2 contains 9,172 full documents. Each document in this collection has at least one table that spans multiple pages or multiple columns within a page, as shown above. In this example, two tables in the document are highlighted: an isolated one in red and a single table split over two columns in blue.
Cropped
Single
Full
Tables
Pages
Documents
Samples
136k
468k
9,172
Tables
136k
548k
25k
Non-English/Multilingual
348
682
34
Pages
—
468k
137k
Two-column pages
—
37%
30%
Table 3: Dataset statistics for PubTables-v2. The data is organized into three collections by surrounding context: cropped tables, single pages, and full documents.
Figure 5: Qualitative failure analysis for table PMC10436220_table_0 (34 rows × 5 columns). Each cell in panels (b)–(d) shows the predicted text content, with the content of predicted spanning cells repeated at each spanned grid location; red-bordered cells contain errors, with background shading proportional to error magnitude (white = perfect match, red = complete mismatch). SmolDocling exhibits correct structure with minor character-level errors (e.g., α→ a). DeepSeek-OCR exhibits errors recognizing the true rowspans of the spanning cells in the first column. Gemini 3.1 Pro fails to recognize the spanning cells altogether, segmenting them into multiple content and blank cells.
Figure 6: Two samples from the cross-page table continuation task. Left ( + ): a positive pair where the table continues onto the next page. Right ( – ): a negative pair. Both pairs are from the same document.
Model
Recall
Precision
F1
AUC
ResNet-50
0.986
0.973
0.979
0.991
ViT-B-16
0.995
0.987
0.991
0.996
Table 7: Cross-page table continuation classification. Results for models trained on PubTables-v2’s continuation prediction task.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 7: This figure shows 5 of the longer multi-page tables from PubTables-v2, with the bounding box for each table highlighted in a particular color on each contiguous page it spans.
Figure 8: Total number of objects per class in the PubTables-v2 Single Pages collection. As can be observed, there is a wide variance in class frequencies. Overall, rotated objects are much rarer than non-rotated objects.
Figure 9: Annotation layers shown separately for four pages. Top row: original page images. Middle row: page-level object associations—table ( blue ), caption ( teal ), and footer ( orange ) bounding boxes with arrows indicating which caption/footer belongs to which table. Bottom row: cell-level structure—individual cell bounding boxes colored by type: regular cells ( gray-blue ), spanning cells ( magenta ), and projected row headers ( gold ). This decomposition highlights the two complementary annotation layers in PubTables-v2: document-level object relationships and fine-grained table structure.
Collection
No. of Parsing Errors
No. of Test Samples
SmolDocling
GraniteDocling
Cropped Tables
16
43
7,680
Single Pages
958
35
7,680
Appendix
Table 9: DocTags Parsing Errors. This table shows the total numbers of parsing errors for DocTags outputs of SmolDocling and GraniteDocling . We use the docling-core Python package (version 2.48.4 ) to parse the DocTags model outputs.
Figure 10: Qualitative failure analysis for table PMC10747712_table_3 from the Cropped Tables collection (17 rows × 19 columns). Cell-level GriTS precision grids are shown for four models, with red intensity indicating content error magnitude. Errors exhibited include numeric quantity failures, additional rows, additional columns, cell misalignments, and spanning cell failures. Note that SmolDocling (b) drops some columns entirely, which contributes to a lower score but cannot be visualized here.
Model
Samples
Recall
Precision
F1
ResNet-50
250
0.908
0.809
0.856
1,000
0.973
0.921
0.946
4,000
0.986
0.958
0.972
15,830
0.986
0.973
0.979
ViT-B-16
250
0.902
0.824
0.862
1,000
0.983
0.914
0.947
Appendix
Table 15: Cross-page table continuation classification ablation study. In this experiment, we show the impact that the number of unique training samples has on performance.
Figure 11: This figure shows one of the examples that both models from the cross-page table continuation prediction experiment fail on. Both models erroneously predict that the last table on the first page (left) continues onto the second page (right).
Figure 12: This figure shows the shortest document in the test set of the PubTables-v2 Full Documents collection, which has only two pages.
Figure 13: Ranking stability vs. test set size on Single Pages (1,000 bootstrap subsamples per N ). Three progressively stricter metrics are shown: all pairs —the fraction of (pair, subsample) combinations that agree with the full-set ordering; adjacent pairs —restricted to consecutively-ranked models, where score gaps are smallest; and complete ranking —the fraction of subsamples that preserve the entire model ordering simultaneously. While the all-pairs metric appears reassuring ( ∼ 90% even at N=63 ), adjacent-pair agreement drops to 67% and complete ranking collapses to < 1%, revealing that small test sets cannot reliably distinguish close models.