jdsouza's picture
Rename public release to Scientific Table Reconstruction
bc19893 verified
|
Raw History Blame Contribute Delete
5.85 kB
metadata
license: cc-by-4.0
task_categories:
  - image-to-text
  - image-text-to-text
language:
  - en
pretty_name: Scientific Table Reconstruction
size_categories:
  - n<1K
tags:
  - scientific-tables
  - table-reconstruction
  - ocr
  - tabulusbench
  - rms
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/*.parquet

Scientific Table Reconstruction

Image-to-table reconstruction across 610 physical table crops from 253 scientific papers in five domains, released by SciKnowOrg, derived from TabulusBench Step 2. This release includes P253. It exposes the Step 2 task independently: reconstruct each provided table image as rows of cell strings.

Leaderboard · Evaluator and submission guide · Tabulus · Source archive

Load

from datasets import load_dataset
ds = load_dataset("SciKnowOrg/Scientific-Table-Reconstruction", revision="v1.0.1", split="test")
example = ds[0]
image = example["image"]          # decoded PIL image
rows = example["gold_rows"]       # list of rows; each cell is a string

One public test split. Each row has exactly these five fields:

Field Type Meaning
domain string Existing broad scientific domain
paper_id string Existing paper identifier, e.g. P253
table_id string Physical crop identifier, e.g. page_011_table_001
image Image Original crop bytes, embedded in Parquet
gold_rows list[list[string]] Existing source gold CSV parsed as rows of strings

The pair (paper_id, table_id) uniquely identifies a sample. IDs and domain are for routing and reporting; model inputs are image-only. Text, empty strings, multiline cells and source row order are preserved. One gold CSV has ragged row widths; no source cells have been repaired or rewritten. Header/row-label interpretations belong to the evaluator, not the release schema. No PDFs or downstream step annotations are included.

Population

Domain Papers Table crops
Agriculture Food And Environmental Systems 50 107
Biomedicine And Health 50 127
Computer Science AI And Data Science 50 122
Energy Materials Chemical Sciences 53 159
Engineering Robotics And Built Infrastructure 50 95
Total 253 610

Each physical fragment is evaluated independently. Multiple crops may belong to one logical table, or continue across pages. Paper IDs are retained so researchers can avoid splitting crops from the same paper across any derived train/test partition. This release defines no training split and no logical-table merging.

Evaluation

Both RMS (Tabulus's DePlot-derived Relative Mapping Similarity) and Exact Mapping Match report macro and micro precision, recall and F1, for twelve aggregate measures. Evaluation mode is physical-table and input mode is image-only. Default ranking is macro RMS F1.

Macro averages per-table measures over all 610 crops, including failures. Micro pools matched credit and gold/predicted mapping counts first. Exact mapping compares identical normalized keys and values; it is not raw CSV equality or positional cell accuracy. Each metric considers original and transposed prediction orientations. Full thresholds, empty-set conventions, normalization and reproducibility details are in the linked evaluator guide.

This is an open-gold benchmark, without a hidden test set. Gold must not be used for inference or tuning; submissions disclose potential training overlap. Complete predictions and provenance are required, and maintainers rescore before accepting a leaderboard entry.

The inherited mapping representation treats the first row/column as labels and collapses duplicate combined keys. The audit finds 285 crops with fewer parsed mappings than their rectangular body-cell count (duplicate labels and/or parser effects), one crop with a literal mapping delimiter, one ragged source grid, and one zero-mapping gold crop. Flags can overlap. These are limitations of the frozen v1 scoring representation; source gold remains unchanged. Aggregate gold mapping denominator: 24,234. Gold nonempty cells: 45,069. All scores must identify this protocol, version and fingerprint.

Version and provenance

  • Dataset version: v1.0.1
  • Protocol: tabulusbench-step2-v1
  • Content fingerprint: f026daa52fe17c24e66068f474cb7fef5f309d3e688ad72bc07e0c412040784e

manifest.json records the source metric hashes, frozen evaluator hashes, domain counts, scoring caveat counts and Parquet hash. Gold rows and image bytes were verified against the local canonical catalog. Changes to gold, population or mapping rules require a new version. The independently distributed Step 6 DOI annotations and Step 7 exports are outside this dataset.

Source, attribution and license

Derived from the TabulusBench dataset whose Zenodo archive declares CC BY 4.0: https://zenodo.org/records/20230340. Preserve that attribution and the original scientific-publication attribution when redistributing material. Paper IDs retain the link to the canonical TabulusBench paper catalog.

The metric is adapted from Google Research DePlot under Apache 2.0, and the standalone evaluation wrapper uses the Tabulus software's MIT license. Software licenses do not replace dataset/source-image attribution.

For citation, use the source archive's Cite panel and identify this Step 2 dataset version and content fingerprint. This card does not introduce an unverified manuscript citation.