sciencialab/grobid-training

GROBID Training-Data Source PDFs

Source PDF documents collected for the GROBID training corpora, paired with a manifest describing which GROBID sequence-labelling models each PDF has hand-annotated training data for.

• 2,484 PDFs, ~3.43 GB total (median 441 KB, range 6 KB – 50 MB)
• 1,588 PDFs are referenced by at least one model's training data
• 896 PDFs were collected but are not (yet) annotated for any model

Files

File	Purpose
*.pdf	the source documents
metadata.csv / metadata.parquet	one row per PDF; keyed on file_name

metadata.csv and metadata.parquet hold identical data. The file_name column is the join key that the 🤗 Datasets folder loader uses to attach metadata to each file.

Usage

from datasets import load_dataset

# metadata + resolved PDF paths
ds = load_dataset("<this-folder-or-repo>")
print(ds["train"][0])

# or just the manifest, no file loading
import pandas as pd
meta = pd.read_parquet("metadata.parquet")
segmentation_docs = meta[meta["segmentation"] == 1]["file_name"].tolist()

Metadata schema

Column	Type	Description
file_name	string	Exact PDF filename; the loader join key.
doc_id	string	Normalized id — filename with .pdf and decorations (.full, -keywords-segmentation, -header, …) stripped.
source	string	Provenance of the file: original-collection, biorxiv, arxiv, arxiv-old, or doi.
size_bytes	int64	File size in bytes.
has_training_data	bool	True if the PDF is referenced by ≥1 model's training data.
n_models	int64	Number of models with training data for this PDF.
models_list	string	;-separated list of those model names.
segmentation … shorttext	int64	One 0/1 flag per model (see below).

Per-model flag columns

segmentation, header, fulltext, figure, table, reference-segmenter, citation, affiliation-address, date, name-header, name-citation, shorttext — each is 1 when the PDF has hand-annotated training data for that GROBID model, else 0.

Annotated-PDF counts per model:

Model	PDFs	Model	PDFs
segmentation	881	reference-segmenter	90
header	822	fulltext	83
name-header	315	date	44
affiliation-address	245	figure	28
citation	166	table	21
shorttext	13	name-citation	9

Provenance

source	Count	Notes
original-collection	2,003	PDFs already present in the collection.
doi	162	Open-access copies resolved via Unpaywall.
grobid-training-data	151	Source PDFs taken from the private grobid-training-data batch repo to fill referenced-but-missing ids.
biorxiv	102	Downloaded from bioRxiv.
arxiv	65	Downloaded from arXiv (modern ids).
arxiv-old	1	Old-style arXiv id (hep-th/0506081v3).

Non-original-collection PDFs were missing from the original collection but referenced by existing training data; their model coverage was backfilled from the missing-list annotations, so has_training_data and the per-model flags are populated for them too. The grobid-training-data PDFs were only added when their id matched a referenced-but-missing training id (not the batches' broader un-annotated collection pool).

Notes & caveats

• Only open-access documents were downloadable. 195 paywalled DOIs and 2 IP-blocked bioRxiv items could not be retrieved and are absent here.
• doc_id is best-effort normalization; a few filenames carry irregular decorations. Always join on file_name, not doc_id.
• This dataset ships the source PDFs and a coverage manifest — the actual TEI/.training. annotation files live in the GROBID repository under grobid-trainer/resources/dataset/<model>/corpus/.

License

The manifest is released under CC-BY-4.0. Individual PDFs retain the license of their original publisher/preprint server; consult each source before redistribution.