metadata extraction (title, authors, affiliations, abstract, …),
bibliographical reference extraction, parsing and citation-context identification, and
full-text body structuring.

Each collection pairs published PDFs with a gold/reference XML (JATS or NLM) that GROBID's output is scored against. In total the dataset contains ~5,927 articles across four collections (~15 GB).

Collections

PMC_sample_1943 — a PubMedCentral gold-standard dataset compiled by Alexandru Constantin, ~1.5 GB, 1943 articles from 1943 different journals (a 2011 snapshot of the PMC Open Access Subset). Each article ships a PDF, an NLM XML file, and a PDFX v1.9 conversion. See PMC_sample_1943/__README__ for provenance.
biorxiv-10k-test-2000 — 2000 bioRxiv preprint articles, originally compiled and published by Daniel Ecer on Zenodo, ~5.4 GB. Each article ships a PDF and the corresponding reference NLM file (created by bioRxiv). The NLM files were further systematically reviewed by the GROBID team and annotated with additional markup for the four statement types described in Statement-level annotations below.
PLOS_1000 — 1000 PLOS articles randomly selected from the full PLOS Open Access collection, ~1.3 GB. Each article ships the published PDF and the corresponding publisher JATS XML file. A PLOS_1000/download.sh script records the original PDF source URLs.
eLife_984 — 984 articles randomly selected from the eLife open collection on GitHub, ~4.5 GB. Each article ships the published PDF, the publisher JATS XML file, and the eLife public HTML file (included as a bonus; not used in scoring), all in their latest version.

Dataset structure

Every collection is a directory of one sub-directory per article. The file naming differs slightly per source:

grobid-evaluation-hf/
├── PMC_sample_1943/
│   ├── __README__                       # provenance (Constantin, PDFX)
│   └── <Journal_Year_.../>              # one folder per article
│       ├── <id>.pdf                     # published PDF
│       ├── <id>.nxml                    # NLM reference XML (gold)
│       └── <id>.pdfx.xml                # PDFX v1.9 conversion
├── biorxiv-10k-test-2000/
│   └── <id>v<n>/                        # e.g. 002485v1
│       ├── <id>v<n>.pdf
│       └── <id>v<n>.xml                 # NLM reference XML (Grobid-team enriched)
├── PLOS_1000/
│   ├── download.sh                      # original PDF download URLs
│   └── <id>/                            # e.g. pbio.0020278
│       ├── <id>.pdf
│       └── journal.<id>.xml             # JATS reference XML (gold)
└── eLife_984/
    └── <id>/                            # e.g. 00013
        ├── elife-<id>-v<n>.pdf
        ├── elife-<id>-v<n>.xml          # JATS reference XML (gold)
        └── elife-<id>-v<n>.html         # bonus, not used in scoring

Statement-level annotations

Beyond the publisher-provided body structure, the GROBID end-to-end evaluation scores four "statement" fields extracted from the article back matter. Where the source XML tags them inconsistently (or not at all), the GROBID team harmonised / added the markup — most notably on the bioRxiv NLM files (systematically reviewed and annotated) and via a disambiguating type="contribution" attribute on PLOS JATS, where fn-type="con" is reused for both conflict-of-interest and author-contribution footnotes.

Statement	Reference-XML markup recognised	Notes
Data / code availability	`//sec[@sec-type="availability"]`, `//p[@content-type="availability"]`, `//sec[@sec-type="data-availability"]` (eLife), `//custom-meta[@id="data-availability"]` (PLOS)
Funding	`//sec[@sec-type="funding"]`, `//p[@content-type="funding"]`, `//funding-statement` (PLOS)
Conflict / competing interests	`//fn[@fn-type="conflict"]`, `//fn-group[@content-type="competing-interest"]`, `//fn[@fn-type="COI"]`, `//sec[@sec-type="COI-statement"]`
Author contributions	`//fn[@type="contribution"]` (team-added disambiguation), `//fn-group[@content-type="author-contribution"]`, `//sec[@sec-type="contribution"]`	a.k.a. CRediT / author contributions

These correspond to GROBID's availability_stmt, funding_stmt, conflict_stmt and contribution_stmt evaluation fields. Not every article contains every statement; the markup is present only where the corresponding statement exists in the source article.

How to use

This dataset is a file tree of PDFs and reference XML — it is not a tabular dataset, so use huggingface_hub to download it rather than datasets.load_dataset:

from huggingface_hub import snapshot_download

path = snapshot_download(
    repo_id="<owner>/grobid-evaluation",   # replace with the actual repo id
    repo_type="dataset",
    allow_patterns=["PLOS_1000/*"],        # optional: fetch one collection
)

Point GROBID's end-to-end evaluation at the downloaded collection following https://grobid.readthedocs.io/en/latest/End-to-end-evaluation/.

Licensing

These resources are originally published under the CC-BY license. The additional annotations contributed by the GROBID team (data/code-availability and funding markup on the bioRxiv set) are similarly released under CC-BY 4.0.

We thank NIH/PMC, bioRxiv, PLOS and eLife for making these resources Open Access and reusable.

Citation

If you use this dataset, please cite the GROBID end-to-end benchmark archive:

@dataset{grobid_end_to_end_evaluation,
  title  = {GROBID end-to-end evaluation datasets},
  author = {GROBID team},
  year   = {2023},
  url    = {https://zenodo.org/record/7708580},
  note   = {https://grobid.readthedocs.io/en/latest/End-to-end-evaluation/}
}

The PMC sample additionally derives from:

@inproceedings{constantin2013pdfx,
  title     = {PDFX: fully-automated PDF-to-XML conversion of scientific literature},
  author    = {Constantin, Alexandru and Pettifer, Steve and Voronkov, Andrei},
  booktitle = {Proceedings of the 2013 ACM Symposium on Document Engineering},
  year      = {2013},
  publisher = {ACM}
}

Downloads last month: 679

Total file size:

16.2 GB

Collection including sciencialab/grobid-evaluation

Grobid

Collection

6 items • Updated Apr 8 • 1