pubpub/pubtables-qa_v2

PubTables-QA v2: A Multi-Page Document Table QA Benchmark

PubTables-QA is a benchmark for evaluating vision-language models on document-level table question answering over multi-page scientific papers. Questions require reading tables that span several pages, cross-referencing multiple tables, and reasoning jointly over tables and the surrounding body text.

Dataset Summary

	Count
QA pairs	1,871
Documents	296
Page images	4,552
Median pages per document	14
Categories	10

The current release contains 1,871 questions. It contains 627 human-authored and 1,244 generated questions.

Source

All documents come from PubTables-v2 (PMC Open Access scientific papers), test split only.

Structure

pubtables-qa_v2/
├── README.md
├── data/
│   └── test.jsonl
├── images/
│   ├── PMC10239856/
│   │   ├── page_0.jpg
│   │   └── ...
│   └── ...
└── metadata/
    ├── annotations.jsonl
    └── authorship.jsonl

Data Fields (data/test.jsonl)

Field	Type	Description
qid	string	Unique question identifier
question	string	Natural language question
answer	string	Gold answer
images	list[string]	All page images of the document (relative paths)
evidence_pages_printed	list[int] | null	Printed journal page labels for the evidence pages, when available
evidence_pages	list[string]	For the pre-restoration 2026-09-24 subset, pages needed to answer: the pages of the cells the answer uses; all pages of a table when the question scans the whole table (max/min, counts, sums, ranks, trends, structure, continuation); plus, for Document-level items, the page(s) of the supporting body-text sentence. Items built from table-structure certificates (qid _hd_, _xp_, _cl_) list the pages their certificate requires. annotation.evidence_pages_by_source in the metadata splits Document-level pages into table and body-text pages.
doc_id	string	PMC document identifier
case_name	string | null	Fine-grained question type. Present on the original items and on augmented items added from 2026-09-23 on (generator family name); null on earlier augmented ones, which are typed by category instead
level	string	Reasoning level
category	string	Question category
source	string	core (original) or augmented
num_pages	int	Total pages in the document

Reasoning Levels

Level	Count	Description
Document Level	322	Body text and tables must be read together
Table Level	368	Structure, span, or whole-table patterns
Cell/Column Level	1,181	Locating and computing over specific cells

Categories

Document Level

Category	Count
Cross-section Synthesis	67
Document Structure	73
Text-Table Reasoning	182

Table Level

Category	Count
Cross-page Continuation	84
Cross-table Structure Match	68
Table Layout	48
Table-wide Reasoning	168

Cell/Column Level

Category	Count
Cross-table Retrieval-Computation	350
Multi-hop	279
Single-table Retrieval-Computation	552

Metadata (metadata/annotations.jsonl)

annotations.jsonl and authorship.jsonl each contain exactly the same 1,871 question IDs as the test split.

Provenance for every question — the structured facts the question was built from.

Field	Type	Description
qid	string	Matches data/test.jsonl
question	string	Question text
answer	string	Gold answer
annotation	object	Structured evidence (see below)

annotation contains the following fields when available:

Key	Description
task	Generation task name, when recorded
tables	Referenced tables, when recorded: table_id (position in the document), printed_label (the label printed on the page, e.g. Table 3), caption, page_indices, n_rows (data rows, header excluded), n_cols, column_headers. Filled from the PubTables-v2 table annotation matched to evidence_bboxes; a few tables absent from that annotation carry only page_indices.
evidence_bboxes	page_id, page_index, table_id, bbox in page-image pixels. Covers the fragments of the referenced tables that lie on the question's evidence_pages.
evidence_values	Supporting cells: table_id, row_label, column_header, value. For structure questions (row/column counts, continuation) the entries describe the table itself: row_label is the printed table label, column_header names the attribute (printed_label, caption, page_indices, n_pages, n_data_rows, n_rows_including_header, n_cols, column_headers) and value its value; the first such entry also carries the full table dump in extra (type: table_metadata, incl. cells as Markdown).
body_text	Supporting sentences from the body text (Document-level items): page_id, text, and — where the sentence was located on the page image — page_index and bbox (page-image pixels, union of the matched words) with location_source = word_match (automatic) or human_review (placed or corrected by a reviewer). The 2026-09-24 subset carries located page IDs; restored historical items retain the available source locations.
evidence_pages_by_source	(Document-level items only) table: the table pages needed to answer; body_text: the pages carrying the supporting body-text sentence(s). evidence_pages in data/test.jsonl is the union of the two.
reasoning	Reasoning steps recorded at generation time
extras	Task-specific fields kept verbatim. continuation_labels (Table-/Cell-level items) lists continuation markers such as “Table 1 (Continued)” with their page_id; these are table evidence, not body text.
evidence_values_value	(core Cell/Column-level items only, 523 of 890) The minimal set of cells actually used to derive the answer, one entry per cell: table_id, table_index (PubTables-v2 index), row_idx, col_idx, row_label, column_header, value, page_id, page_index, bbox (page-image pixels; null when the same value occurs at several positions and could not be disambiguated), role (answer, operand, left/right, numerator/denominator, condition, step, candidate), source (how the cell was located). Unlike evidence_values, which for core items lists the whole referenced table, this field is limited to the cells the question needs.
evidence_scans	Column scans the answer depends on (e.g., maximum, count or comparison over a column): table_id, column_header, page_indices, what.
evidence_captions	Captions or footnotes used as evidence: page_index, table_id, text.
evidence_layout	Page-layout observations used as evidence (e.g., where a table starts or the printed page number): page_index, fact.
evidence_values_value_meta	Provenance of the field above: method (rule_operand_join from the generation-time operand fields, or gpt5_grid_pick from GPT-5 selecting coordinates on the indexed table grid), status (complete/partial), arith_ok (whether the operand cells reproduce the gold answer, when checkable), flags. The remaining 367 core Cell/Column-level items do not carry this field. Of the 523 annotated items, 511 have status complete and 12 partial (current release).

For every question, the evidence fields describe the current question and answer.

Usage

import json
from PIL import Image

with open("data/test.jsonl") as f:
    data = [json.loads(line) for line in f]

sample = data[0]
print(sample["question"], sample["answer"])

pages = [Image.open(p) for p in sample["images"]]          # whole document
evidence = [Image.open(f"images/{sample['doc_id']}/page_{p.rsplit('_page_', 1)[1]}.jpg")
            for p in sample["evidence_pages"]]              # evidence pages only

Citation

@article{pubtablesqa2026,
  title={PubTables-QA: A Multi-Page Document Table QA Benchmark},
  year={2026}
}

License

CC-BY-4.0. Source document images come from PubTables-v2 (PMC Open Access).