fairdataihub/poster-sentry-training-data

PosterSentry Training Data

Human-validated training dataset for PosterSentry, the multimodal scientific poster classifier used in the posters.science quality control pipeline. Each row holds the text extracted from the document's first page and the exact feature values the released model was trained on.

Developed by the FAIR Data Innovations Hub at the California Medical Innovations Institute (CalMI²).

Version

Version	Date	Notes
1.1.0	2026-10-28	Every document is released by its depositor under CC0, CC-BY, or an equivalent permissive license, with an attribution sheet for all of them. First-page text is extracted with the poster2json layout engine. Adds the raw extracted text and the feature values used in training (text embedding, text score, visual and structural features).
1.0.0	2026-08-18	Human-validated labels from a three-reviewer survey and a blinded adjudication review, with document identifiers, DOIs, and label provenance.

The earlier unversioned release (April 2026) used heuristic repository labels and is superseded.

Dataset Description

Text and features from real scientific documents, zero synthetic data. Every sample comes from an actual PDF downloaded from Zenodo or Figshare as part of the posters.science corpus, and every label was validated by people. No PDFs or page images are redistributed.

The corpus holds 3,235 documents: 1,605 posters and 1,630 non-posters (near-balanced by construction of the candidate pools, not by resampling).

Labeling

1. Survey: Three reviewers independently classified 3,414 candidate documents (poster, non-poster, or unsure) at survey.posters.science, casting 10,242 votes with an inter-rater Krippendorff's alpha of 0.79.
2. Adjudication: The 420 documents without a unanimous panel (362 decided two to one, 58 exact ties) were adjudicated in a blinded review: each document image was re-examined in randomized order without access to the panel votes or any model output.
3. Deduplication: 172 near-duplicate documents (matching normalized 300-character text prefixes) and 7 documents with unavailable PDFs were removed, leaving the 3,235 documents published here.

Non-posters include multi-page papers, conference proceedings, abstract books, newsletters, flyers, slide decks, and other documents labeled as "posters" in repository metadata.

Licenses and attribution

Every document in this dataset was released by its depositor under CC0, CC-BY, or an equivalent permissive license (a public-domain mark, MIT, or Apache-2.0): CC-BY 4.0 (3,143), CC0 1.0 (75), MIT (9), CC-BY 3.0 (3), Apache-2.0 (2), U.S. Public Domain (2), Public Domain Mark 1.0 (1). The title, creators, year, DOI, and license of every source record are listed in PosterSentry_CC-BY_credits.csv (also as PosterSentry_CC-BY_credits.xlsx), which is the attribution these licenses require; the same sheet is archived with the Zenodo record (10.5281/zenodo.20221832).

Files

File	Description	Rows
data/train.parquet	One row per training document: identifiers, license, label, raw and cleaned first-page text, and every feature value the model was trained on	3,235
data/poster_sentry_train.ndjson	The same documents without the feature columns (text, labels, identifiers)	3,235
data/poster_sentry_features.csv	The feature values alone, one column per feature (emb_000 to emb_511 for the text embedding)	3,235
PosterSentry_CC-BY_credits.csv / .xlsx	Attribution for every source document (title, creators, year, repository, DOI, license)	3,414

Fields

Field	Description
id	Survey document identifier (matches the survey, adjudication, and label files in the Zenodo record)
doi, doi_link	DOI of the source repository record (Zenodo or Figshare) and its resolvable link
source	zenodo or figshare
license	License of the source record (for example CC-BY-4.0 or CC0-1.0)
label	poster or non_poster (human-validated)
label_source	unanimous_panel (all non-unsure votes identical) or adjudicated (settled in the blinded review)
text_raw	First-page text as extracted by the poster2json layout engine on pdfplumber (headings marked with ## ); for 3 documents where it returns no text, pdfplumber's default extraction
text	text_raw whitespace-normalized and truncated to 4,000 characters: the text the model embeds
extraction	poster2json or pdfplumber_default
embedding	The 512-dimensional model2vec (minishlab/potion-base-32M) embedding of text, L2-normalized
text_score_oof	The stage-one text score of the document from five-fold out-of-fold prediction, as fed to the final classifier when the released head was fit
img_width ... high_contrast_ratio	The 15 visual features of the rendered first page
page_count ... size_per_page_kb	The 15 structural features from PDF layout analysis

The final classifier reads text_score followed by the 15 visual and 15 structural features, in the column order above. Feature definitions are in Table 1 of the paper and in the poster-sentry source.

Label Distribution

Label	Count
poster	1,605
non_poster	1,630

By provenance: 2,822 documents carry a unanimous panel label and 413 an adjudicated label. Every row carries a DOI.

Related Resources

Resource	Link
PosterSentry model	fairdataihub/poster-sentry
poster-sentry	GitHub
poster-sentry-training	GitHub
poster-sentry-evaluation-paper-code	GitHub
Survey votes, adjudication, labels, credits (archive)	Zenodo 10.5281/zenodo.20221832
Llama-3.1-8B-Poster-Extraction	fairdataihub/Llama-3.1-8B-Poster-Extraction
poster2json library	PyPI · GitHub
poster-json-schema	GitHub
Platform	posters.science

Usage

from datasets import load_dataset
import numpy as np

ds = load_dataset("fairdataihub/poster-sentry-training-data", split="train")
row = ds[0]
print(row["id"], row["label"], row["license"], row["text"][:80])
X_text = np.array(ds["embedding"], dtype="float32")   # (n, 512)

Use for PubGuard doc_type training

The poster texts in this dataset are also used by PubGuard to train its poster document-type classification head.

Citation

@dataset{poster_sentry_data_2026,
  title = {PosterSentry Training Data: Scientific Poster Text Corpus},
  author = {O'Neill, Jamey and Portillo, Dorian and Zeinali, Nahid and Soundarajan, Sanjay and Blake, Gerard and Sarin, Parth and Buttrick, Adam and Patel, Bhavesh},
  year = {2026},
  version = {1.1.0},
  doi = {10.57967/hf/8865},
  url = {https://huggingface.co/datasets/fairdataihub/poster-sentry-training-data},
  note = {Part of the posters.science initiative}
}

License

The dataset is released under CC-BY 4.0. Each source document remains under its own license, listed per document in the license column and in the attribution sheet. The PosterSentry model and code are released under the MIT license.

Acknowledgments

• FAIR Data Innovations Hub at California Medical Innovations Institute (CalMI²)
• posters.science platform
• HuggingFace for dataset hosting infrastructure
• Funded by The Navigation Fund (10.71707/rk36-9x79), "Poster Sharing and Discovery Made Easy"