SEBENARNYA-Fact

SEBENARNYA-Fact is a Malay-language fact-checking dataset for evidence retrieval derived exclusively from publicly accessible fact-checking articles published by SEBENARNYA.MY.

The dataset accompanies the manuscript:

SEBENARNYA-Fact: A Malay Fact-Checking Dataset for Evidence Retrieval

The finalized version contains 4,039 records. Each retained record represents a claim–evidence pair with source metadata, publication information, normalized verdict metadata, and traceable source URLs.

Dataset summary

SEBENARNYA-Fact is designed primarily for article-level evidence retrieval in automated fact-checking.

Each retained claim is treated as a retrieval query, while its paired evidence passage is treated as the relevant document. The normalized fact-check verdict is metadata and is not used to define retrieval relevance.

The dataset was constructed through article discovery, content extraction, normalization, deduplication, structural quality assurance, label adjudication, and final snapshot validation.

Final adjudication

The structurally cleaned snapshot contained 4,045 records. During label normalization, 141 records were flagged for review.

Following adjudication:

• 135 records were retained with finalized normalized labels;
• 6 records were excluded because they did not contain sufficient substantive evidence to form reliable claim–evidence pairs;
• the final dataset contains 4,039 records; and
• 0 unresolved needs_review labels remain.

Dataset structure

The release contains:

data/SEBENARNYA_Fact_Dataset_v1.0.csv
SEBENARNYA-Fact-HuggingFace-v1.0.ipynb

The CSV preserves the finalized dataset schema used by the accompanying reproducibility pipeline. The notebook provides an executable companion workflow for using the Hugging Face release.

Main fields

The dataset includes the following types of fields:

• record_id — stable record identifier
• claim_text — checked claim or issue used as the retrieval query
• evidence_text — extracted fact-checking evidence/explanation
• evidence_summary — concise evidence summary
• label_original — source-side or extracted label information
• label_normalized — finalized normalized label
• source — source organization
• source_url — original article URL
• canonical_url — canonicalized article URL
• language — detected language
• domain — topical/domain label where available
• date_published — publication date
• date_collected — collection date
• retrieval_doc_id — retrieval document identifier
• additional collection/provenance fields retained from the construction pipeline

The exact schema should be treated as authoritative from the CSV header.

Label distribution

Normalized label	Count	Percentage
False	2,279	56.42%
Clarification	1,264	31.29%
Warning	394	9.75%
Misleading	85	2.10%
True	17	0.42%
Total	4,039	100.00%

No needs_review labels remain in the finalized release.

Retrieval benchmark formulation

For the accompanying evidence-retrieval benchmark:

• each claim_text is a query;
• its paired evidence document is the single relevant document;
• the evidence document is constructed from the complete evidence passage when sufficiently populated, otherwise from the evidence summary;
• query/document relevance mapping is deterministic and row-aligned;
• evaluation uses nDCG@10, Recall@10, and MRR@10.

The executable benchmark-construction and retrieval pipeline is released separately in the associated code repository.

Baseline results

The finalized 4,039-record benchmark produced the following results:

Model	nDCG@10 (%)	Recall@10 (%)	MRR@10 (%)
BM25	88.68	94.31	86.84
DistilUSE-multilingual	54.53	68.68	50.06
LaBSE	70.04	82.00	66.22
multilingual-E5	83.79	91.66	81.25
BGE-M3	88.33	94.11	86.45
Qwen3-Embedding-0.6B	86.62	92.94	84.57
Jina v5 Nano	78.15	86.16	75.56
BM25 + BGE-M3	90.22	95.22	88.58
BM25 + Qwen3-Embedding-0.6B	90.04	95.07	88.39

Robustness diagnostics

A lexical-overlap analysis was conducted using Jaccard similarity between each claim and its paired evidence.

• Mean Jaccard overlap: 0.1187
• Median Jaccard overlap: 0.1042
• Lowest-overlap quartile threshold: ≤ 0.0702
• Lowest-overlap subset: 1,014 queries

Performance on the lowest-overlap subset:

Model	nDCG@10 (%)	Recall@10 (%)	MRR@10 (%)
BM25	73.59	83.04	70.55
BGE-M3	76.86	84.42	74.41
Qwen3-Embedding-0.6B	75.51	83.14	73.01

These diagnostics are intended to complement the primary benchmark rather than to claim that the current dataset represents open-world fact checking.

Intended uses

SEBENARNYA-Fact may support research on:

• Malay-language evidence retrieval;
• automated fact-checking pipelines;
• lexical, dense, and hybrid information retrieval;
• multilingual retrieval;
• Malaysian misinformation research;
• retrieval robustness and error analysis.

Out-of-scope uses and limitations

The dataset should not be treated as:

• a complete representation of misinformation in Malaysia;
• an open-world fact-checking benchmark;
• a balanced veracity-classification dataset;
• a substitute for the original SEBENARNYA.MY articles;
• a guarantee that all extracted text is free from extraction or normalization errors.

Important limitations include:

1. Single-source construction. The dataset is derived from SEBENARNYA.MY only.
2. Same-article claim–evidence pairing. Claims and evidence originate from the same fact-checking article, which may make retrieval easier than external-claim retrieval.
3. Label imbalance. False and clarification records dominate the final distribution.
4. Article-level evidence. The benchmark does not provide sentence-level evidence annotation.
5. Derived source content. Some fields contain text extracted or derived from third-party fact-checking articles.
6. Temporal snapshot. The dataset reflects the collected website snapshot and may not represent later changes to the source site.

Source attribution and data-use notice

SEBENARNYA-Fact is derived from publicly accessible content published by SEBENARNYA.MY. Source URLs are retained to support attribution and traceability.

The dataset creators do not claim ownership over third-party article content originating from SEBENARNYA.MY. The license: other metadata is used intentionally because redistribution and reuse of source-derived text may be subject to the source website's applicable terms, copyright, and other rights.

Users are responsible for ensuring that their use of the dataset complies with applicable law and the terms governing the original source content.

See DATA_USE_NOTICE.md for the release notice included with this repository.

Privacy and responsible use

The dataset is intended for research on public fact-checking and information retrieval. It should not be used to infer sensitive personal attributes, target individuals, or make high-stakes decisions about people.

Source text may mention people, organizations, public incidents, or claims as part of fact-checking coverage. Users should consult the original source article and exercise appropriate care when interpreting such content.

Versioning and maintenance

This release is:

SEBENARNYA-Fact v1.0 — 4,039 records

Future releases should use explicit version numbers and document any additions, removals, corrections, or changes to the benchmark construction.

If the dataset changes, retrieval experiments should be rerun so that reported benchmark results remain synchronized with the released dataset snapshot.

Authors

Ibrahim Mohd Jamadi

Faculty of Artificial Intelligence, Universiti Teknologi Malaysia (UTM), Kuala Lumpur, Malaysia
Email: ibrahim98@graduate.utm.my
ORCID: 0009-0007-7571-3354

Syahid Anuar

Faculty of Artificial Intelligence, Universiti Teknologi Malaysia (UTM), Kuala Lumpur, Malaysia
Email: syahid.anuar@utm.my
ORCID: 0000-0001-7869-0257

Corresponding author: Ibrahim Mohd Jamadi

Citation

If you use this dataset, please cite the associated paper:

Ibrahim Mohd Jamadi and Syahid Anuar, “SEBENARNYA-Fact: A Malay Fact-Checking Dataset for Evidence Retrieval,” 2026.

The final conference/proceedings citation and DOI should replace this provisional citation once available.

Code

The reproducible dataset-construction and retrieval pipeline is released separately through the associated GitHub repository:

https://github.com/ibrahimjamadi98/SEBENARNYA-Fact