PubTables-1M: Table Extraction Benchmark

• PubTables-1M is a large-scale dataset that provides comprehensive supervision for table extraction tasks by aligning PDF coordinates with XML-derived table structures.
• It supports three key subtasks—table detection, structure recognition, and functional analysis—using precise cell geometry and header semantics to overcome previous annotation ambiguities.
• Its canonicalization process and rigorous quality filtering significantly reduce oversegmentation, thereby enhancing the reliability of model evaluation and enabling diverse downstream applications.

PubTables-1M is a large-scale dataset for table extraction from unstructured documents, introduced to provide comprehensive and less ambiguous supervision for scientific-document tables. It was developed from PubMed Central Open Access (PMCOA) articles and defines table extraction as comprising table detection, table structure recognition, and functional analysis. The dataset contains nearly one million tables, is grounded in PDF coordinates rather than only raster images, and was designed to address limitations of earlier datasets, especially incomplete spatial annotation, weak cell-location labels, missing header semantics, and markup-derived inconsistency caused by oversegmentation (Smock et al., 2021).

1. Definition and scope

PubTables-1M supports three subtasks. Table detection (TD) locates tables on a page. Table structure recognition (TSR) recovers rows, columns, and cells, including spanning cells. Functional analysis (FA) identifies semantic roles, especially headers versus data regions (Smock et al., 2021).

The dataset is built from PMCOA scientific articles, where each article is available both as a PDF and as an XML document containing HTML table markup. This pairing is central: the XML provides semantic table structure, while the PDF provides visual presentation. Because the annotations are anchored in PDF coordinates, PubTables-1M supports multiple input modalities, including image-based, PDF/text-based, and potentially multimodal approaches (Smock et al., 2021).

In scale, the dataset reports 947,642 tables for TSR. For table detection, it includes fully annotated pages containing tables, with 460,589 training pages, 57,591 validation pages, and 57,125 test pages. For TSR, the document-level split is 758,849 training tables, 94,959 validation tables, and 93,834 test tables, using an 80/10/10 split at the document level to avoid leakage across tables from the same paper (Smock et al., 2021).

The dataset is explicitly table-centric rather than page-contextualized. Later work has treated this as a strength for cropped-table TSR and as a limitation for unified page understanding. For example, CTE reuses PubTables-1M as a table-annotation backbone and embeds it into full-page context by merging it with PubLayNet, while POTATR is positioned as a later extension of the PubTables/TATR line toward page-level extraction on PubTables-v2 rather than on PubTables-1M itself (Gemelli et al., 2023, Smock et al., 8 Jun 2026).

2. Annotation design and represented structure

PubTables-1M was designed to be more comprehensive than earlier large-scale table datasets. Its annotations include table bounding boxes, explicit row and column bounding boxes, cell topology, spanning cells, full cell locations including blank cells, and header information such as column headers and projected row headers (Smock et al., 2021).

The dataset distinguishes between two cell-localization concepts. A text cell bounding box is the union of character boxes for the text in a cell. A grid cell bounding box is the full geometric cell region derived from row and column intersections. This distinction matters because blank cells have no text extent but still have valid grid-cell geometry. PubTables-1M therefore supports supervision not only for cell content and topology but also for whitespace-sensitive cell geometry and blank-cell prediction (Smock et al., 2021).

For joint TSR and FA, the dataset is used with six object classes:

Class
table
table column
table row
table column header
table projected row header
table spanning cell

An implicit seventh class, table grid cell, is formed by the intersection of rows and columns (Smock et al., 2021).

The paper also defines projected row headers (PRHs) as rows in the body area that act as section labels spanning across the table’s logical structure. The rule used is: for each row, if the row is not in the column header and has exactly one non-blank cell that occupies the first column, then label it a projected row header (Smock et al., 2021).

Later datasets and models preserve and reinterpret these annotations. CTE states that it inherits from PubTables-1M the classes row, column, table header, projected header, table cell, and grid cell, then converts those table-region annotations into token classes and token links for graph-ready page representations (Gemelli et al., 2023). This suggests that PubTables-1M’s annotation granularity has been sufficiently rich to support not only object detection formulations but also token/link reformulations.

3. Construction pipeline and PDF-grounded alignment

The annotation pipeline begins from PMCOA XML and PDF pairs. XML markup encodes semantic table structure but does not directly identify the corresponding PDF coordinates. To recover that correspondence, the authors process each PDF into a sequence of characters with bounding boxes, extract the character sequence from the HTML table markup, and align the two sequences with the Needleman–Wunsch algorithm (Smock et al., 2021).

After alignment, each text-containing cell is assigned a text cell bounding box defined as the union of the matched character boxes. Higher-level annotations are then inferred geometrically. The table bounding box is defined as the union of all text cell bounding boxes. A row bounding box spans the full table width, with vertical extent determined by the union of text cells for all cells whose starting row or ending row is that row. A column bounding box spans the full table height, with horizontal extent determined by the union of text cells for all cells whose starting column or ending column is that column (Smock et al., 2021).

The grid cell bounding box is then defined as the union of the bounding boxes of its rows intersected with the union of the bounding boxes of its columns. This construction yields a full geometric region even for blank cells (Smock et al., 2021).

This PDF-grounded design has downstream consequences. The original paper emphasizes that multiple modalities can be derived from the annotated PDF data (Smock et al., 2021). Later work confirms the practical value of this grounding. CTE uses PubTables-1M annotations together with text extracted from PDFs using PyMuPDF, assigning token labels according to enclosing annotated regions and creating links among tokens belonging to the same row, column, or grid cell (Gemelli et al., 2023). A plausible implication is that the original PDF-coordinate anchoring of PubTables-1M made such later reinterpretations substantially easier.

4. Canonicalization and the oversegmentation problem

A central contribution of PubTables-1M is its treatment of oversegmentation, defined as the case where a cell that should logically be a single spanning header cell is split into multiple smaller grid cells in the annotation. The paper argues that this creates contradictory supervision during training and underestimates model performance during evaluation, especially for complex tables with hierarchical headers (Smock et al., 2021).

The canonicalization procedure is intended to ensure that each table has a unique, unambiguous structure interpretation. The paper motivates this procedure with several assumptions: consistency with the Wang model of tabular structure, unique indexing by keys, a requirement that internal nodes have at least two children, and an assumption that oversegmentation preserves visual appearance (Smock et al., 2021).

The algorithm first extends header annotations. It splits blank spanning cells into blank grid cells, may add the first row to the column header if the first row starts with a blank cell, may iteratively extend the column header until every column in the column header has at least one complete cell spanning only that column, labels projected row headers using the one-non-blank-cell-in-first-column rule, and may extend the row-header region when cells in the first column below the column header are spanning or blank (Smock et al., 2021).

It then performs recursive merges under specific conditions: column-header cells are merged with adjacent cells above or below that span the exact same columns; blank cells may be merged upward or downward under specified all-blank conditions; projected row headers are merged across the full row; and row-header cells may be recursively merged with adjacent blank cells below (Smock et al., 2021).

The paper reports that canonicalization modifies 34.7% of all tables, or 65.8% of complex tables. It also states that 52.7% of tables are complex after canonicalization, compared with 40.1% before canonicalization (Smock et al., 2021).

Evidence for reduced ambiguity is given through projected row header oversegmentation statistics. Counting a PRH as oversegmented if its row contains a blank cell, the paper reports 0% PRH oversegmentation for PubTables-1M, compared with 15.79% for SciTSR, 58.65% for PubTabNet, and 98.87% for FinTabNet (Smock et al., 2021). This is not presented as a complete measure of all oversegmentation, but it supports the claim that the dataset removes a major source of annotation inconsistency.

5. Quality control, evaluation methodology, and baseline results

Because manual review at PubTables-1M scale is infeasible, the dataset is filtered automatically. Tables are discarded if rendered markup yields overlapping rows or columns. For each cell, the paper compares original XML text with PDF text extracted from the corresponding grid-cell box and discards tables whose average normalized edit distance exceeds 0.05. For each word inside the table boundary, it computes overlap with the most-overlapping grid cell and discards tables whose average is below 0.9. It also removes tables with more than 100 objects, with less than 0.1% of tables discarded as outliers (Smock et al., 2021).

Evaluation spans both object-detection metrics and table-structure metrics. For TD and joint TSR+FA object detection, the paper uses COCO-style AP, AP$_{50}$, AP$_{75}$, and AR (Smock et al., 2021). For TSR, it reports table content accuracy, adjacent-cell content F-score, and GriTS variants. The paper gives the GriTS formula as

$\mathrm{GriTS}_f(\mathbf{A}, \mathbf{B}) = \frac{2 \cdot \sum_{i,j} f(\tilde{\mathbf{A}_{i,j}, \tilde{\mathbf{B}_{i,j})} {|\mathbf{A}| + |\mathbf{B}|}$

with variants $\mathrm{GriTS}_{\mathrm{Top}}$, $\mathrm{GriTS}_{\mathrm{Cont}}$, and $\mathrm{GriTS}_{\mathrm{Loc}}$ for topology, content, and location recognition respectively (Smock et al., 2021).

The paper evaluates standard object-detection architectures without task-specific customization: DETR and Faster R-CNN (Smock et al., 2021). For TD, DETR achieves AP = 0.966, AP50 = 0.995, AP75 = 0.988, AR = 0.981, while Faster R-CNN achieves AP = 0.825, AP50 = 0.985, AP75 = 0.927, AR = 0.866 (Smock et al., 2021).

For joint TSR + FA object detection, DETR achieves AP = 0.912, AP50 = 0.971, AP75 = 0.948, AR = 0.942, compared with Faster R-CNN at AP = 0.722, AP50 = 0.815, AP75 = 0.785, AR = 0.762 (Smock et al., 2021).

For TSR evaluation on canonical data, DETR trained and tested on canonical labels reaches, for all tables, $Acc_{Cont} = 0.8138$, $GriTS_{Top} = 0.9845$, $GriTS_{Cont} = 0.9846$, $GriTS_{Loc} = 0.9781$, and $_{75}$0. On complex tables, the same model reaches $_{75}$1 and $_{75}$2 (Smock et al., 2021).

The canonicalization ablation is among the dataset’s most consequential empirical findings. DETR trained and tested on non-canonical data achieves $_{75}$3 and $_{75}$4 for all tables, while the same model evaluated on canonical test data drops to $_{75}$5 and $_{75}$6 (Smock et al., 2021). The paper interprets this as showing both better supervision and more reliable evaluation from canonical labels.

6. Role in later research and benchmark evolution

PubTables-1M has continued to function as both a benchmark and a source of supervision in later work. In "Latent Diffusion for Guided Document Table Generation," PubTables-1M is used as the primary real dataset for training a synthetic table-image generator, as the source of row and column annotations used to create conditioning masks, and as the benchmark for testing whether those synthetic images are useful for downstream row/column structure recognition (Hamdani et al., 2024).

That work states that PubTables-1M contains 575,305 annotated document pages for table detection and 947,642 fully annotated table crops for structure recognition and functional analysis, and uses the split of 758,849 training, 93,834 testing, and 94,959 validation table images for structure-oriented experiments (Hamdani et al., 2024). It trains a latent diffusion model on 450k samples from the PubTables dataset, generates 250k synthetic annotated table images, and trains YOLOv5 only on those synthetic images. Evaluated on the PubTables-1M test set of 93,834 images, that synthetic-only model reaches $_{75}$7 and $_{75}$8, compared with $_{75}$9 and $\mathrm{GriTS}_f(\mathbf{A}, \mathbf{B}) = \frac{2 \cdot \sum_{i,j} f(\tilde{\mathbf{A}_{i,j}, \tilde{\mathbf{B}_{i,j})} {|\mathbf{A}| + |\mathbf{B}|}$0 for YOLOv5 trained on real PubTables data (Hamdani et al., 2024). The paper frames this as evidence that PubTables-1M supervision can be converted into useful synthetic, pre-annotated training data.

FastTab uses PubTables-1M as a direct benchmark for single-table or tightly cropped table TSR. It follows the official split of 758,849 training, 94,959 validation, and 93,834 test tables, and evaluates only GriTS$\mathrm{GriTS}_f(\mathbf{A}, \mathbf{B}) = \frac{2 \cdot \sum_{i,j} f(\tilde{\mathbf{A}_{i,j}, \tilde{\mathbf{B}_{i,j})} {|\mathbf{A}| + |\mathbf{B}|}$1 on PubTables-1M (Hamdi et al., 21 May 2026). In the reported comparison, FastTab achieves 98.27 GriTS$\mathrm{GriTS}_f(\mathbf{A}, \mathbf{B}) = \frac{2 \cdot \sum_{i,j} f(\tilde{\mathbf{A}_{i,j}, \tilde{\mathbf{B}_{i,j})} {|\mathbf{A}| + |\mathbf{B}|}$2 at 39.50 FPS, compared with 86.16 for Faster R-CNN, 98.45 for DETR, 97.62 for DETR-NC, and 99.22 for VAST (Hamdi et al., 21 May 2026). This positions PubTables-1M as a continuing reference benchmark for topology-oriented TSR.

CTE recontextualizes PubTables-1M by merging a filtered overlapping subset with PubLayNet. In that setting, PubTables-1M contributes table-specific supervision, including row, column, header, projected header, table cell, and grid cell signals, which are then projected into token labels and token links for the task of Contextualized Table Extraction (Gemelli et al., 2023). CTE is much smaller, at 75k pages and 35k tables, because it keeps only overlap with PubLayNet and discards misaligned or incomplete pages (Gemelli et al., 2023). This later use underscores that PubTables-1M is not merely a benchmark but also an annotation substrate for newer task formulations.

POTATR represents a different evolutionary path. It does not benchmark on PubTables-1M directly, but it is presented as extending TATR, the DETR-based model associated with the PubTables-1M line, from cropped-table TSR to page-level extraction on PubTables-v2 (Smock et al., 8 Jun 2026). The paper explicitly situates PubTables-1M as the earlier benchmark centered on cropped-table structure recognition, while PubTables-v2 adds page-level and full-document contexts (Smock et al., 8 Jun 2026). This suggests a benchmark-family progression rather than a replacement.

7. Significance, limitations, and interpretive context

PubTables-1M is best characterized as both a scale contribution and a data-quality intervention. Its distinctive claims are not limited to the size of 947,642 tables, but include PDF-grounded spatial supervision, explicit row and column localization, representation of blank cells, header and projected-row-header annotation, automated quality filtering, and canonicalization intended to remove oversegmentation and enforce a unique structure interpretation (Smock et al., 2021).

Its importance for model design is evident in the strong performance of generic object detectors trained on the dataset. The original paper argues that transformer-based object detection models can achieve excellent results for TD, TSR, and FA without special customization (Smock et al., 2021). Later research reinforces the centrality of its annotation scheme: latent diffusion methods treat PubTables-1M row and column annotations as controllable generation conditions (Hamdani et al., 2024); grid-centric recognizers such as FastTab evaluate against its topology benchmark (Hamdi et al., 21 May 2026); and page-level contextual datasets such as CTE inherit and reinterpret its table-annotation vocabulary (Gemelli et al., 2023).

Several limitations are also explicit. The dataset is derived from PMCOA scientific articles, so its construction assumptions and canonicalization heuristics are PMCOA-specific (Smock et al., 2021). The original paper states that multi-page tables are out of scope and that full row-header inference beyond projected row headers is also outside scope (Smock et al., 2021). Later work makes related distinctions: PubTables-1M remains table-centric, whereas CTE is page-contextualized (Gemelli et al., 2023), and PubTables-v2 extends the benchmark family toward page-level and document-level extraction (Smock et al., 8 Jun 2026).

A common misconception is to treat PubTables-1M as simply a larger markup-derived table dataset. The evidence reported in the original paper does not support that reduction. The dataset’s central premise is that better ground truth materially improves both learning and evaluation reliability, and the canonical versus non-canonical experiments are presented precisely to show that annotation quality, not merely dataset size, changes observed model performance (Smock et al., 2021). A plausible implication is that PubTables-1M’s long-term influence comes as much from its formalization of what counts as complete and canonical table supervision as from its scale.