daniB2112/fake-news-dataset

Fake News Dataset

A large-scale, multi-source English-language dataset for binary fake-news detection. The release contains ~9.25M curated and deduplicated samples intended primarily for training robust text-classification models on long-form news articles and headlines.

Key facts

• Size: ~9.25 million samples (train split)
• File provided: train.csv (~4.42 GB)
• Labels: binary label_binary — 0 = FAKE, 1 = REAL
• Language: English
• Primary task: Binary text classification / fact-checking
• Curator: Arko007
• Dataset landing page: https://huggingface.co/datasets/Arko007/fake-news-dataset

Quick start

1. Download train.csv from the dataset repo or HF dataset page.
2. Stream with the Hugging Face datasets library or load as Parquet/CSV locally.
3. When training, use stratified sampling or class weights due to heavy class imbalance (FAKE ≈ 3%).

Example (load CSV with datasets):

from datasets import load_dataset

ds = load_dataset("csv", data_files="train.csv", split="train")
print(ds.column_names)
print(ds[0])

Dataset Description

This dataset aggregates multiple public fake-news corpora, curated publisher content, and Kaggle collections into a single, standardized CSV suitable for large-scale model training. It was assembled to support domain-adaptive pretraining and large-scale fine-tuning (for example, the reference checkpoint Arko007/fake-news-roberta-5M).

Motivation

• Provide a large, diverse training corpus for detecting misinformation in news articles.
• Support multi-stage transfer learning (news-adaptive pretraining → task fine-tuning → specialized models such as political fact-checkers).

Composition

• Combined and deduplicated sources totaling ~9.25M rows.
• Representative class distribution (from the 5M subset used for analysis):
  • FAKE: ~2.8%
  • REAL: ~97.2%

Note on provenance

• To avoid metadata inconsistencies, this public release contains the two canonical fields (text, label_binary) in the primary train.csv.
• Per-record provenance (original source identifiers) is not included in the primary CSV for this release. Aggregated source statistics and provenance mapping files (if available) are provided in the repository under provenance/ or can be requested from the maintainer. This decision avoids exposing third-party metadata that may carry additional redistribution constraints.

Dataset Structure

Files in this release

• train.csv — main CSV file with the fields below (primary artifact; ~4.42 GB)

Primary columns

• text (string): article body, headline, or statement text
• label_binary (int): 0 = FAKE, 1 = REAL

Optional supplemental artifacts (in repo)

• provenance statistics and aggregation summaries (CSV/TSV)
• preprocessing and merging scripts
• checksums for distributed artifacts

Splits

• train: single split included with ~9.25M rows.
• No official validation/test splits are included by default. See "Recommended Splits" below.

Dataset Creation

Curation steps

1. Collected multiple public datasets (ISOT, Kaggle fake-news datasets, COVID-19 misinformation corpora, political fake-news collections, scientific claim datasets).
2. Standardized fields to a unified schema (text, label_binary).
3. Performed duplicate and near-duplicate detection and removal across sources.
4. Cleaned text (unicode normalization, whitespace normalization, basic HTML artifact removal). URLs and certain boilerplate tokens were removed/normalized during preprocessing.
5. Mapped source-specific labels to binary labels: FAKE (0) and REAL (1). Multi-class truthfulness labels from some sources were conservatively mapped to binary assignments during mapping.
6. Assembled final train.csv and recorded aggregated source statistics.

Preprocessing details

• Lowercasing and optional punctuation normalization used for analysis; raw release text retains original casing by default (users may re-normalize).
• Minimal PII redaction in the general pipeline — users deploying models in production should audit and remove sensitive content as required by their policies.

License & provenance

• This repository declares Apache-2.0. Upstream source licenses vary; maintainers must ensure that redistribution under Apache-2.0 is compatible with upstream terms and preserve required notices when applicable.

Uses

Recommended

• Training transformer-based classifiers (RoBERTa, DeBERTa, etc.) for article-level fake-news detection.
• Domain-adaptive pretraining for news and misinformation tasks.
• Baselines and research into imbalance mitigation, calibration, and domain transfer.

Worked examples / reference

• RoBERTa-base fine-tuned on a 5M subset (Arko007/fake-news-roberta-5M) achieved a validation accuracy of ~99.28% on in-domain news tasks.
• Transfer to LIAR (political statements) via further fine-tuning achieved ~71% accuracy for short-statement classification.

Not recommended

• Legal judgments of truth or presenting automated outputs as authoritative.
• Out-of-domain short-text tasks (consider a LIAR- or claims-focused dataset for that use).

Recommended Splits & Training Advice

Suggested default split (if you need validation/test)

• Stratified 80/10/10 split by label_binary (preserve class proportion).
  • Train: 80%
  • Validation: 10%
  • Test: 10%

Because FAKE is rare (~3%), use one or more of these strategies:

• Class weighting in the loss (e.g., inverse-frequency weighting)
• Oversampling FAKE class or synthetic augmentation for FAKE examples
• Hard negative mining or curriculum learning
• Evaluate with metrics robust to imbalance (F1, precision/recall per class, AUPRC)

For temporal generalization tests

• Consider chronological splits (train on older news, validate/test on more recent months/years) to simulate real-world deployment shifts.

Bias, Risks, and Limitations

Known biases

• Source bias: ISOT and other major components can create publisher/style cues that models may exploit rather than learning content-level signals.
• Class imbalance: Extremely skewed toward REAL; naive accuracy is a poor metric.
• Domain & geographic bias: Over-representation of US/Western sources and political topics in some source subsets.
• Temporal bias: Data distribution may reflect particular time periods and not future misinformation strategies.

Risks

• High in-domain accuracy does not guarantee safe performance in the wild; adversarial and out-of-domain fake news can bypass models trained on these sources.
• Models may unduly rely on publisher/domain signals, leading to false positives/negatives when applied to new sources.
• Redistribution and legal risk for copyrighted publisher content — verify rights.

Limitations

• English-only dataset; not suitable for multilingual detection tasks.
• Binary labels remove nuance—no fine-grained truthfulness scale.
• Per-record provenance is not included in the primary train.csv release (see notes above).

Mitigation recommendations

• Evaluate on diverse and out-of-domain benchmarks (e.g., LIAR, PolitiFact corpora).
• Use human-in-the-loop verification for high-stakes applications.
• Report per-class precision/recall and calibration; prefer F1 and AUPRC over raw accuracy.
• Maintain provenance and be prepared to remove items if upstream rights owners request it.

Distribution & Storage

• Primary artifact: train.csv (~4.42 GB)
• Prefer streaming/Parquet or chunked CSV loading for training at scale.
• Repository contains checksums (when available) and recommended ingestion scripts.

Contact & Maintainer

• Maintainer: Anamitra-Sarkar (GitHub: https://github.com/Anamitra-Sarkar)
• Hugging Face dataset page: https://huggingface.co/datasets/Arko007/fake-news-dataset
• For licensing or provenance concerns, open an issue on the dataset repo or contact the maintainer via the GitHub profile.

Citation

If you use this dataset, please cite both the curated dataset and the original sources.

Curated dataset citation

@dataset{arko007_fakenews,
  title = {Fake News Dataset: A Large-Scale Multi-Source Collection},
  author = {Arko007},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/Arko007/fake-news-dataset}
}

Representative source citations

@article{ahmed2017isot,
  title={ISOT Fake News Dataset},
  author={Ahmed, Hadeer and Traore, Issa and Saad, Sherif},
  year={2017}
}

@inproceedings{wang2017liar,
  title={"Liar, Liar Pants on Fire": A New Benchmark Dataset for Fake News Detection},
  author={Wang, William Yang},
  booktitle={ACL},
  year={2017}
}

Changelog

• v1.0 — corrected metadata and filenames; primary release includes train.csv (~9.25M rows) — 2025-10-18
• Notes in this version:
  • Replaced num_rows with num_examples to be compatible with the datasets library split schema.
  • Updated top-level license to Apache-2.0 and adjusted license review notes.
  • Ensured filename references match the actual artifact train.csv.
  • Added Quick Start and recommended stratified splits.