BeliefSim Dataset Release

This folder contains an anonymized, Hugging Face-ready data package for BeliefSim: Towards Belief-Driven Simulation of Demographic Misinformation Susceptibility.

Content Note

This dataset studies real misinformation and rumor examples. Some claims may contain offensive, stigmatizing, or otherwise harmful wording. Examples are included for research transparency and should be handled with care.

Files

• data/claims.csv: claim text from PANDORA and MIST-1, with source labels and gold labels where available.
• data/judgments.csv: anonymized participant-level claim judgments.
• data/evaluation_instances.csv: target claim judgments paired with two same-participant observed belief judgments.
• data/wvs_group_priors.csv: WVS-derived group belief distributions by demographic axis.
• data/dataset_summary.json: generated counts and privacy notes.
• scripts/prepare_release.py: reproducible script used to regenerate the release files from local raw sources.

Current Generated Counts

• Claims: 156 total, including 56 PANDORA and 100 MIST-1 claims.
• Judgments: 40,964 total, including 64 PANDORA and 40,900 MIST-1 judgments.
• Evaluation instances: 40,135 total, including 53 PANDORA and 40,082 MIST-1 instances.
• WVS group prior rows: 976.

These counts are generated from the local files available in this workspace at preparation time.

Anonymization

The release script removes direct identifiers and sensitive survey metadata before writing the public CSV files. In particular, it drops response IDs, Prolific IDs, IP addresses, latitude/longitude, raw timestamps, and free-text notes/comments. Participant identifiers are replaced with deterministic salted SHA-256 hashes. Demographic fields are kept only as broad categories such as gender, age group, education bucket, and living-area group where available.

Before public upload, verify that the upstream PANDORA, MIST-1, and WVS data-use terms permit redistribution of the fields included here. If redistribution is not permitted for any source, upload only the scripts/metadata and provide instructions for authorized users to reconstruct the dataset locally.

Loading

from datasets import load_dataset

claims = load_dataset("MichiganNLP/beliefsim", data_files="data/claims.csv")
judgments = load_dataset("MichiganNLP/beliefsim", data_files="data/judgments.csv")
instances = load_dataset("MichiganNLP/beliefsim", data_files="data/evaluation_instances.csv")
wvs_priors = load_dataset("MichiganNLP/beliefsim", data_files="data/wvs_group_priors.csv")

Citation

@article{borah2026belief,
  title={Belief-Sim: Towards Belief-Driven Simulation of Demographic Misinformation Susceptibility},
  author={Borah, Angana and Khan, Zohaib and Mihalcea, Rada and P{\'e}rez-Rosas, Ver{\'o}nica},
  journal={arXiv preprint arXiv:2603.03585},
  year={2026}
}

Questions or collaboration inquiries about misinformation research can be directed to anganab@umich.edu.