Hugging Face's logo

Datasets:
SciKnowOrg
/
tabulus-gold

Languages:
English
Tags:
scientific-literature
table-extraction
document-understanding
information-extraction
bibliography
citation-matching
License:
Dataset card Files
xet
 Community
tabulus-gold / README.md
jdsouza's picture
jdsouza
Update README.md
70462f8 verified
preview code
|
Raw History Blame Contribute Delete
8.37 kB
metadata 
pretty_name: Tabulus-Gold
license: other
tags:
  - scientific-literature
  - table-extraction
  - document-understanding
  - information-extraction
  - bibliography
  - citation-matching
  - scholarly-communication
  - atomic-layer-deposition
viewer: false
language:
  - en

Tabulus-Gold

Tabulus-Gold provides gold-standard data for evaluating the output quality of the seven-Step Tabulus pipeline for reference-aware scientific review-table digitization.

Software: https://github.com/sciknoworg/tabulus
Documentation: https://tabulus.readthedocs.io/en/latest/

The release contains controlled gold data for two review papers from the atomic layer deposition (ALD) literature:

  • Miikkulainen et al. (2013) — Crystallinity of inorganic films grown by atomic layer deposition: Overview and general trends. Journal of Applied Physics, 113(2). DOI: 10.1063/1.4757907
  • Puurunen (2005) — Surface chemistry of atomic layer deposition: A case study for the trimethylaluminum/water process. Journal of Applied Physics, 97(12), 121301. DOI: 10.1063/1.1940727

The data are organized by paper and pipeline Step. Each Step directory contains the authoritative target data—or, for deterministic downstream processing, the expected derived output—against which Tabulus predictions can be evaluated.

What is Tabulus?

Tabulus is a modular framework for digitizing scientific review tables into structured, citation-aware data. Its seven-Step workflow covers table localization and reconstruction, identification of reference-bearing tables, bibliography extraction, citation-to-bibliography matching, scholarly-reference resolution, and export of resolved structured tables.

Dataset structure 

tabulus-gold/
├── README.md
├── Miikkulainen_2013/
│   ├── README.md
│   ├── step1_table_localization/
│   ├── step2_table_reconstruction/
│   ├── step3_reference_table_classification/
│   ├── step4_bibliography/
│   ├── step5_reference_matching/
│   ├── step6_reference_resolution/
│   └── step7_resolved_tables/
└── Puurunen_2005/
    ├── README.md
    ├── step1_table_localization/
    ├── step2_table_reconstruction/
    ├── step3_reference_table_classification/
    ├── step4_bibliography/
    ├── step5_reference_matching/
    ├── step6_reference_resolution/
    └── step7_resolved_tables/

Gold-standard contracts by pipeline Step

Step Tabulus task Gold-standard target
Step 1 — Table localization Identify physical table fragments in the source PDF and preserve the information required to associate fragments belonging to the same logical table. gold.json describes the curated physical table fragments and available localization/continuation metadata. Canonical table-crop images provide the visual table regions used as the controlled input to Step 2.
Step 2 — Table reconstruction Convert each canonical table crop into a structured table while preserving its rows, columns, and cell contents. Each curated physical table has an authoritative gold.csv representing the expected reconstructed table. These files define the target table structure and content independently of the reconstruction model or adapter used.
Step 3 — Reference-table classification Determine whether a reconstructed table contains an explicit literature-reference column and should therefore continue through the reference-processing branch of Tabulus. gold.json assigns the authoritative reference-table classification to each curated Step 2 table. The gold set contains both positive and negative examples where available.
Step 4 — Bibliography extraction Recover the ordered bibliography of the source paper so that table citations can subsequently be linked to bibliography positions. gold.json contains the authoritative ordered bibliography. Bibliography position is important because Step 5 links table citations to these ordered entries.
Step 5 — Reference matching Detect citation expressions occurring in reference-bearing table cells and map them to the corresponding positions in the Step 4 bibliography. gold.json records the annotated citation cells and their gold bibliography indices. A cell may link to one or multiple bibliography entries.
Step 6 — Scholarly reference resolution Resolve bibliography entries linked from the tables to canonical scholarly identities, such as DOI, title, authors, year, and venue. Resolution is performed once at paper level rather than independently for every table occurrence. gold.json contains source-derived scholarly identities for the controlled resolution set. These records provide the authoritative target metadata for evaluating resolution quality.
Step 7 — Resolved-table export Join the reconstructed table content, Step 5 bibliography links, and Step 6 scholarly identities into the final reference-aware structured tables. The Step 7 gold.csv files are deterministically derived expected outputs. They combine authoritative upstream gold data and represent the final table content and scholarly-reference fields expected from a correct end-to-end join/export.

Controlled evaluation data

Across the two papers, Tabulus-Gold contains:

  • 65 curated physical tables
  • 57 tables with explicit reference columns
  • 8 tables without explicit reference columns
  • 2,016 annotated citation cells
  • 5,963 citation-to-bibliography links
  • 3,596 ordered bibliography entries
  • 3,276 source-derived Step 6 scholarly-reference records
  • 57 Step 7 resolved physical-table gold CSVs

These counts correspond to different parts of the seven-Step evaluation workflow and therefore do not share a single common evaluation denominator.

For example, Step 3 is evaluated over both reference-bearing and non-reference-bearing curated tables, whereas Steps 5 and 7 operate on the reference-bearing subset.

Relationship to Tabulus-Runs

Tabulus-Gold contains the authoritative annotations and expected outputs used for controlled quality evaluation.

The separate Tabulus-Runs release contains recorded experimental outputs from the ALD and area-selective deposition (ASD) experiments used in the associated Tabulus study.

Keeping the two resources separate distinguishes the evaluation targets from the system outputs being evaluated.

Citation

When using Tabulus-Gold, please cite the Zenodo dataset record. The final dataset DOI and recommended citation will be added here when the record is published.

If you use the Tabulus software, please also cite the software repository:

@software{tabulus,
  author = {D'Souza, Jennifer and Rumleanschi, Vladimir},
  title = {Tabulus: End-to-End Scientific Review Table Digitization Pipeline},
  url = {https://github.com/sciknoworg/tabulus},
  year = {2026}
}

Tabulus was first developed in the following master's thesis:

@mastersthesis{rumleanschi2026tabulus,
  author = {Rumleanschi, Vladimir},
  title = {Tabulus: AI-Assisted Tabular Knowledge Extraction and Semantic Integration Framework for Scientific Literature},
  year = {2026},
  doi = {10.15488/21618},
  url = {https://doi.org/10.15488/21618}
}

License

The original annotations, structured gold-standard data, and other dataset content created for Tabulus-Gold are licensed under the Creative Commons Attribution-ShareAlike 4.0 International License (CC BY-SA 4.0).

You are free to share and adapt these materials for any purpose, provided that appropriate credit is given to the creators and that any adapted material is distributed under the same or a compatible license.

Third-party material derived from the source publications, including any table images or table crops, is not covered by this license unless explicitly stated otherwise. Copyright and reuse rights for such material remain with the respective copyright holders.