Jezzarax/pubhealth-converted

PUBHEALTH

PUBHEALTH is a public-health fact-checking dataset introduced in:

Neema Kotonya and Francesca Toni. 2020. Explainable Automated Fact-Checking for Public Health Claims.

This repository is a scriptless Parquet conversion of the original PUBHEALTH TSV files. It is intended to load with the Hugging Face datasets library without requiring deprecated remote dataset scripts.

The deprecated script-based dataset was available at:

https://huggingface.co/datasets/bigbio/pubhealth

The reference scripts used for this conversion were found there. The conversion is needed because current versions of the Hugging Face datasets library no longer support loading external dataset scripts in the old way. See the Hugging Face forum discussion:

https://discuss.huggingface.co/t/dataset-scripts-are-no-longer-supported/163891

Dataset Structure

This upload provides the two configs exposed by the deprecated loader.

pubhealth_source

This config preserves the original source schema:

pubhealth_bigbio_pairs

This config mirrors the deprecated BigBio pairs transformation:

Expected converted files:

data/pubhealth_source/train-00000-of-00001.parquet
data/pubhealth_source/validation-00000-of-00001.parquet
data/pubhealth_source/test-00000-of-00001.parquet
data/pubhealth_bigbio_pairs/train-00000-of-00001.parquet
data/pubhealth_bigbio_pairs/validation-00000-of-00001.parquet
data/pubhealth_bigbio_pairs/test-00000-of-00001.parquet
metadata/conversion_report.json

Conversion

Regenerate the Parquet files from the original raw TSV directory:

uv run convert_pubhealth.py ~/Downloads/PUBHEALTH --output-dir .

The raw directory must contain:

train.tsv
dev.tsv
test.tsv

The conversion script follows the behavior of the deprecated Hugging Face/BigBio loader used as reference:

• —Read tab-separated rows with no CSV quoting.
• —Use the last 9 fields of each row. This drops the extra leading index-like column present in test.tsv.
• —Skip rows with fewer than 9 fields.
• —Skip rows whose label is not one of true, false, unproven, or mixture.
• —Write both pubhealth_source and pubhealth_bigbio_pairs.
• —Write an auditable metadata/conversion_report.json with split counts, skipped row reasons, label counts, and duplicate claim_id counts.

Split Counts

The exact counts depend on the raw files used. After running the converter, see metadata/conversion_report.json for the authoritative counts for this upload.

For the raw files checked during this conversion setup, the old-builder-compatible parser produced:

Loading

from datasets import load_dataset

source = load_dataset("YOUR_NAMESPACE/PUBHEALTH", "pubhealth_source")
pairs = load_dataset("YOUR_NAMESPACE/PUBHEALTH", "pubhealth_bigbio_pairs")

Attribution

If you use this dataset, cite the original PUBHEALTH work:

@article{kotonya2020explainable,
  title={Explainable automated fact-checking for public health claims},
  author={Kotonya, Neema and Toni, Francesca},
  journal={arXiv preprint arXiv:2010.09926},
  year={2020}
}

The deprecated reference loader identified the dataset homepage as:

https://github.com/neemakot/Health-Fact-Checking/tree/master/data

The deprecated Hugging Face/BigBio dataset page used as the script reference was:

https://huggingface.co/datasets/bigbio/pubhealth

License Notes

The deprecated reference loader listed the dataset license as MIT. Before publishing a public reupload, verify the redistribution status of the original raw files and any third-party text included in the fact-check/article fields.

Conversion Notes

This repository intentionally does not include the old pubhealth.py dataset script or BigBio helper script. The goal is to provide the same two dataset views in a scriptless format that current datasets can load directly from Parquet.

This conversion keeps the old loader's two public configs, pubhealth_source and pubhealth_bigbio_pairs, but replaces script execution with explicit Parquet files and dataset-card configs metadata.