JayNightmare's picture
Update README.md
ac955fb verified
|
Raw
History Blame Contribute Delete
11.7 kB
---
pretty_name: FakeNews Master Dataset
size_categories:
- 100K<n<1M
task_categories:
- text-classification
- text-generation
task_ids:
- explanation-generation
tags:
- fake-news-detection
- fact-checking
- claim-verification
- multimodal-metadata
- context-ablation
- research
license: other
language:
- en
multilinguality:
- multilingual
---
# FakeNews Master Dataset
The FakeNews Master Dataset is a merged, cleaned, and deduplicated research corpus for studying misinformation classification under different context conditions. It combines locally available mirrors of ClaimReview, Fakeddit, FakeNewsNet, and MuMiN into one canonical schema while preserving source-dataset provenance.
This dataset is intended for experiments that ask whether open-source models can classify misinformation when given different amounts or qualities of context, such as full context, minimal context, or misleading context.
## Dataset Summary
| Field | Value |
| ------------------------------- | -------------------------------------------------- |
| Total rows before deduplication | 374,738 |
| Total rows after deduplication | 360,348 |
| Source datasets | ClaimReview, Fakeddit, FakeNewsNet, MuMiN |
| Primary task | Binary misinformation classification |
| Internal label convention | `mapped_label: 0 = real`, `mapped_label: 1 = fake` |
| Flat export label convention | `Real: 1 = real`, `Real: 0 = fake` |
| Modalities | Text plus image/news metadata where available |
## Source Composition
| Source dataset | Cleaned rows | Raw rows used | Retention rate | Acquisition notes |
| -------------- | -----------: | ------------: | -------------: | ---------------------------- |
| ClaimReview | 265,884 | 271,083 | 0.9809 | Pinned `2026_06_12` snapshot |
| Fakeddit | 85,577 | 100,000 | 0.8558 | Sample fallback mirror |
| FakeNewsNet | 21,889 | 23,196 | 0.9437 | Local CSV mirror |
| MuMiN | 1,388 | 1,404 | 0.9886 | CSV export fallback |
## Files
The dataset folder includes three export families:
| File pattern | Purpose |
| -------------------------------------- | ----------------------------------------------------------------------------------------------- |
| `test split.{json,jsonl,csv}` | Test split for adapter training and evaluation. |
| `train split.{json,jsonl,csv}` | Training split for adapter training. |
| `validation split.{json,jsonl,csv}` | Validation split for adapter training. |
| `normalized_records.{json,jsonl,csv}` | Deduplicated canonical `UnifiedRecord` corpus for experiments and training. |
| `master_records.{json,jsonl,csv}` | Provenance-rich master records with normalized source fields. |
| `master_flat_records.{json,jsonl,csv}` | Flat analysis export with `id`, `Title`, `Claim`, `news source`, `Real`, and `tweet_num`. |
| `master_dataset_manifest.json` | Build manifest with source counts, acquisition notes, cleaning reports, and coverage summaries. |
| `master_dataset_coverage.json` | Aggregate and per-source coverage report for flat export fields. |
For most Hugging Face `datasets` workflows, start with `normalized_records.jsonl` if you want the full canonical schema, or `master_flat_records.csv` if you want a compact spreadsheet-like table.
## Schema
### Canonical Records
The canonical exports use the repository's `UnifiedRecord` schema. Key fields include:
| Field | Description |
| --------------------- | -------------------------------------------------------------------------- |
| `dataset` | Source dataset name. |
| `sample_id` | Source-aware stable sample identifier. |
| `text` | Primary claim, post, headline, or article text used by the model pipeline. |
| `context_text` | Additional source context when available. |
| `original_label` | Source-specific label value. |
| `original_label_name` | Human-readable source-specific label. |
| `mapped_label` | Internal binary label, where `0 = real` and `1 = fake`. |
| `mapped_label_name` | Internal binary label name, `real` or `fake`. |
| `split` | Dataset split value when assigned. |
| `modality` | Source modality descriptor. |
| `has_image` | Whether image metadata is available. |
| `metadata` | Source-specific provenance and auxiliary fields. |
| `cleaning_notes` | Notes emitted by the cleaning pipeline. |
### Flat Records
The flat export is designed for inspection and analysis tools:
| Field | Description |
| ------------- | ----------------------------------------------------------- |
| `id` | Master dataset row identifier. |
| `Title` | Title/headline when available. |
| `Claim` | Primary claim text. |
| `news source` | Source, publisher, platform, or domain field. |
| `Real` | Flat binary label where `1 = real` and `0 = fake`. |
| `tweet_num` | Tweet/social count when available, mostly from FakeNewsNet. |
## Field Coverage
Aggregate flat-export coverage:
| Field | Non-null rows | Null rate |
| ------------- | ------------: | --------: |
| `Title` | 106,125 | 0.7055 |
| `Claim` | 360,348 | 0.0000 |
| `news source` | 360,348 | 0.0000 |
| `Real` | 360,348 | 0.0000 |
| `tweet_num` | 20,732 | 0.9425 |
High null rates for `Title` and `tweet_num` are expected because not every source dataset provides those fields.
## Loading Examples
### JSONL Canonical Schema
```python
from datasets import load_dataset
dataset = load_dataset(
"json",
data_files="normalized_records.jsonl",
split="train",
)
print(dataset[0])
```
### CSV Flat Schema
```python
from datasets import load_dataset
dataset = load_dataset(
"csv",
data_files="master_flat_records.csv",
split="train",
)
print(dataset.features)
```
### Label Conversion Reminder
The canonical pipeline and flat analysis export use opposite binary conventions:
```python
# Canonical normalized_records: mapped_label 0=real, 1=fake
canonical_label = row["mapped_label"]
# Flat master_flat_records: Real 1=real, 0=fake
is_real = row["Real"]
```
## Dataset Creation Pipeline
![dataset](./etc/dataset.png)
The builder reuses each source-specific loader, applies the shared cleaning pipeline, maps labels into a binary research schema, deduplicates records by stable fingerprints, and writes the canonical, master, and flat exports.
## Intended Uses
This dataset is intended for:
- Research on misinformation and fake-news classification.
- Context-ablation experiments with open-source language models.
- Comparing model behavior across datasets, labels, and context settings.
- Building deterministic train/validation/test splits for adapter fine-tuning.
- Inspecting source coverage and field availability across merged misinformation datasets.
## Out-of-Scope Uses
This dataset should not be used as a sole source of truth for:
- Automated content moderation.
- Real-world fact-checking without human review.
- Legal, medical, election, financial, or safety-critical decisions.
- Ranking people, publishers, communities, or organizations as trustworthy or untrustworthy.
Source labels reflect the conventions, sampling choices, and historical snapshots of the underlying datasets. They are research labels, not universal truth claims.
## Biases, Risks, and Limitations
- The merged corpus is source-imbalanced: ClaimReview accounts for most rows.
- Individual source datasets use different collection methods, label definitions, time periods, languages, and domains.
- Fakeddit and MuMiN are represented through local/sample fallback mirrors in this build.
- Some rows contain little context beyond a claim or title, which affects context-sensitive experiments.
- Image URLs or image metadata may be present for some records, but this master export does not guarantee image availability or long-term URL validity.
- Deduplication reduces repeated content but cannot guarantee complete removal of near-duplicates.
- The flat `Real` label intentionally inverts the internal label convention for analysis compatibility; check labels carefully before training.
## Licensing and Source Terms
This master dataset is an aggregated research artifact derived from multiple upstream datasets. Users are responsible for complying with the terms, licenses, and citation requirements of each original source dataset.
The repository code is MIT licensed, but that does not automatically grant identical rights for every upstream data record. Treat this dataset as `license: other` unless and until each upstream source license is verified for your use case.
## Citation
The project paper is currently being drafted for submission to a peer-reviewed conference.
**BibTeX:**
```bibtex
@misc{compaan_fakenews_master_dataset_2026,
title = {FakeNews Master Dataset for Context-Aware Misinformation Classification},
author = {Compaan, Jay Bell and Peiling, Yi and Charles, Michael J. C.},
year = {2026},
note = {Research dataset artifact; paper in preparation}
}
```
**APA:**
Compaan, J. B., Peiling, Y., & Charles, M. J. C. (2026). *FakeNews Master Dataset for Context-Aware Misinformation Classification*. Research dataset artifact; paper in preparation.
## Authors and Contact
- Dr. Yi Peiling - Kingston University, UK
- Jay Bell Compaan (JayNightmare) - Kingston University, UK
- Michael JC Charles - Kingston University, UK
For questions or comments about this dataset, please contact Dr. Yi Peiling at <yi.peiling@kingston.ac.uk>.
## Reproducibility
The master dataset can be rebuilt from the repository with:
```bash
python src/datasets/fetch_datasets.py
python src/build_master_dataset.py --dataset all --output-dir src/datasets/data/master_dataset
```
Important outputs:
- `master_dataset_manifest.json`
- `master_dataset_coverage.json`
- `normalized_records.jsonl`
- `master_records.jsonl`
- `master_flat_records.csv`
## Dataset in Use
This dataset was used to train the `FakeNews` adapter, which is available on Hugging Face at [JayNightmare/FakeNews](https://huggingface.co/JayNightmare/FakeNews). The adapter was trained on cleaned splits from the master dataset and is intended for research on context-aware misinformation classification.