Datasets:

Syokan
/

DisinfoMM

pretty_name: DisinfoMM
language:
  - en
  - it
  - pt
multilinguality:
  - multilingual
task_categories:
  - image-classification
  - text-classification
size_categories:
  - 10K<n<100K
license: other
tags:
  - disinformation
  - multimodal
  - out-of-context
  - fact-checking
  - evidence
configs:
  - config_name: default
    data_files:
      - split: train
        path: Dataset.csv
dataset_info:
  features:
    - name: Num
      dtype: int64
    - name: Website
      dtype: string
    - name: Date
      dtype: string
    - name: Image
      dtype: string
    - name: Claim
      dtype: string
    - name: Evaluation
      dtype: string
    - name: Label
      dtype: string
    - name: Keywords
      dtype: string
    - name: Tags
      dtype: string
    - name: Article Link
      dtype: string
    - name: Declaration Link
      dtype: string
    - name: Explanation
      dtype: string
    - name: Article sources
      dtype: string
    - name: Process
      dtype: string

DisinfoMM

DisinfoMM is the dataset accompanying “A Multilingual, Multimodal Dataset for Disinformation and Out-of-Context Analysis with Rich Supportive Information” (ICMI 2025, pages 643–651, DOI).

The release intentionally contains the original Dataset.csv as the single authoritative data file. JSON and JSONL files used by individual experiments are derived formats and are generated by the companion GitHub repository; they are not separate versions of the dataset.

Dataset summary

Property	Value
Records	25,752
Snopes / English	16,730
Polígrafo / Portuguese	7,270
Pagella Politica / Italian	1,752
CSV size	112,739,300 bytes
CSV SHA-256	`fa93ec2f4979521932815e9fa2d63df0c72ea1f3cf81e493024d4dcefe66145b`

The paper describes the collection as approximately 25k records. The exact release contains 25,752 rows.

What is—and is not—in the release

Dataset.csv includes the claim text, original verdict, harmonized five-level evaluation, explanation, image URL, fact-check URL, declaration URL, article evidence links, dates, keywords, tags, and source-processing metadata.

Image binaries are not included. The authors’ local archival directories do not contain a complete copy of all images, and source websites retain rights in their media. The Image column is therefore the canonical media reference. The companion repository provides a resumable downloader, but links can expire or change over time.

Columns

Column	Meaning
`Num`	Stable source-snapshot row identifier. Values are unique but not necessarily contiguous.
`Website`	Fact-checking source: `snopes`, `poligrafo`, or `pagella`.
`Date`	Publication date as recorded by the source. Formats reflect the source website.
`Image`	URL of the image associated with the fact-check.
`Claim`	Claim or statement being assessed.
`Evaluation`	Harmonized five-level evaluation produced by the collection pipeline. Historical spelling/capitalization variants are retained.
`Label`	Original source-specific verdict, when available. Pagella rows can leave this field empty because their harmonized evaluation was inferred from article text.
`Keywords`	Source-provided or extracted topic keywords.
`Tags`	Additional tags extracted from important evidence sources.
`Article Link`	URL of the fact-checking article.
`Declaration Link`	URL of the original declaration or claim, when available.
`Explanation`	Explanation/verdict text used as supportive information in the paper.
`Article sources`	Evidence URLs cited by the fact-checking article. Multiple URLs are stored in the historical comma-separated representation.
`Process`	Evidence domains retained by the source-frequency filtering process.

Harmonized labels

The historical Evaluation values contain capitalization and spelling variants such as Mostly true, Moslty true, exagerated or missing details, and Moslty false or misleading. The companion code normalizes them without modifying this archival CSV.

Normalized label	Records	Paper binary mapping
`True`	7,131	authentic (`0`)
`Mostly True`	2,688	authentic (`0`)
`Incomplete`	98	disinformation (`1`)
`Mostly False`	3,429	disinformation (`1`)
`False`	12,404	disinformation (`1`)
`Unknown`	2	excluded until manually reviewed

This binary conversion follows the experiment convention recovered from the archived scripts. Researchers interested in five-level classification should retain the normalized five-level labels instead.

Class balance and experiment subsets

The full CSV is intentionally not balanced: it is the collected dataset snapshot. The paper experiments selected task-specific subsets from this CSV, balanced authentic and disinformation examples, and then created 8:1:1 training/validation/test splits. The paper used approximately 7,000 English samples and 10,000 multilingual samples.

The companion repository provides a deterministic prepare-experiments command that performs this selection and writes both a direct JSONL format and the historical data.json plus annotation JSON format. This keeps the original dataset separate from experimental sampling decisions.

Loading

With pandas:

import pandas as pd

dataset = pd.read_csv("Dataset.csv")
print(dataset.shape)  # (25752, 14)

With Hugging Face Datasets after publication:

from datasets import load_dataset

dataset = load_dataset("Syokan/DisinfoMM", split="train")

CSV cells can contain long explanations and embedded commas. Use a standards-compliant CSV reader rather than splitting lines manually.

Known data-quality conditions

The rows are preserved rather than silently discarded:

35 rows have an empty Claim;
2 rows have an empty Explanation;
2 rows have an unresolved harmonized evaluation;
image URLs are present in all rows, but continued availability is controlled by the source websites;
repeated image or article URLs can be legitimate because a source article may discuss multiple claims or records.

The preparation command reports all exclusions and writes an audit file for every generated experiment subset.

Collection and provenance

The dataset was collected from Snopes, Polígrafo, and Pagella Politica. The released code contains one source adapter per website, label harmonization, evidence-link extraction, and the frequency-based evidence-domain lists described in the paper. New collection runs must be stored separately from this fixed snapshot so that experiments remain reproducible.

Intended use

Appropriate uses include multilingual multimodal disinformation research, out-of-context analysis, evidence-aware training, label-harmonization studies, and cross-domain evaluation. The dataset is not a substitute for professional fact-checking and should not be used to profile individuals or provide medical, legal, or financial advice.

Claims and explanations may contain false, harmful, sensitive, or outdated material. Preserve source attribution and verdict context when presenting examples.

Licensing and third-party rights

Author-created annotations and harmonized metadata are offered under CC BY 4.0. Source articles, quotations, images, videos, trademarks, and linked evidence remain subject to their original owners’ rights and terms and are not relicensed. See DATA_LICENSE.md and THIRD_PARTY.md. This mixed-rights status is why the dataset card uses license: other.

Citation

@inproceedings{cui2025disinfomm,
  author    = {Shuhan Cui and Hanrui Wang and Ching-Chun Chang and Huy H. Nguyen and Isao Echizen},
  title     = {A Multilingual, Multimodal Dataset for Disinformation and Out-of-Context Analysis with Rich Supportive Information},
  booktitle = {Proceedings of the 27th International Conference on Multimodal Interaction},
  pages     = {643--651},
  year      = {2025},
  doi       = {10.1145/3716553.3750813}
}

Contact

For dataset and reproduction questions, contact Shuhan Cui at syokan [at] g.ecc.u-tokyo.ac.jp or open an issue in the GitHub repository.