Dataset Card for CrediBench 1.1

CrediBench is a large-scale, temporal webgraph constituted of web data pulled from Common Crawl.

Dataset Details

Dataset Description

This dataset is composed of monthly slices of large-scale web networks. These webgraphs contain 1+ billion edges, and 45+ million nodes per month. In these webgraphs, the nodes represent a website domain (e.g, google.com) and an edge represents a directed hyperlink relation (e.g, an edge from cbc.ca to reuters.com indicates that a page on cbc.ca's website contains a hyperlink to a reuters.com page). These webgraphs are supplemented with text attributes, partly from Common Crawl and from web scraping, as text features play an important role in misinformation detection. Additionally, we supplement them with credibility scores as made available by Lin et al., to enable supervised and semi-supervised learning as explained in our paper.

Dataset Statistics:

Month	V	E	Min. deg.	Mean deg.	Max. deg.	Leaves (deg. = 1)	Edge Density
October 2024	50,288,479	1,074,971,387	1	42.75	17,112,352	30,278	4.3e-07
November 2024	50,684,724	1,164,563,814	1	45.95	17,328,063	31,388	4.5e-07
December 2024	45,030,252	1,014,523,551	1	45.06	14,719,077	28,857	5.0e-07
January 2025	45,626,949	1,060,163,646	1	46.471	15,398,279	23,130	5.0e-07
February 2025	49,639,664	1,167,748,533	1	47.05	17,078,954	24,430	4.7e-07
March 2025	50,162,733	1,212,826,396	1	48.36	16,691,193	22,629	4.8e-07
April 2025	50,050,221	1,237,519,870	1	49.45	16,679,192	26,639	4.9e-07
May 2025	50,517,253	1,227,682,479	1	48.604	16,771,274	29,973	4.8e-07

Content Embedding:

Domain-level content embeddings are generated using multiple LLM-based embedding models with varying LLM-model sizes and embedding dimensions. The embeddings are intended to support feature initialization for downstream GNN models. For each domain, the textual content is first extracted and then encoded into dense vector representations using the selected embedding model.

The dataset is organized by month under the content_embeddings directory.

Each pickled file stores a dictionary:

{ domain1:[[page_url1, embedding_vector1],[page_url2, embedding_vector2], ...],
  domain2:[[page_url1, embedding_vector1],[page_url2, embedding_vector2], ...],
  ...
}  

Month	Embedding-model	Emb-dim	Total-files-size
October 2024	embeddinggemma-300m	256	30GB
November 2024	embeddinggemma-300m	256	30GB
December 2024	embeddinggemma-300m	256	30GB

Proportion of content

Task	Split	Class	Global Labels	CDB-Oct24 Labels	CDB-Oct24 Has_Content	CDB-Oct24 content(%)	CDB-Nov24 Labels	CDB-Nov24 Has_Content	CDB-Nov24 content(%)	CDB-Dec24 Labels	CDB-Dec24 Has_Content	CDB-Dec24 content(%)
Regression	Train	—	5,164	5,158	2,699	0.52	5,158	2,626	0.51	5,158	2,637	0.50
Regression	Val	—	1,722	1,719	869	0.51	1,719	842	0.49	1,719	841	0.49
Regression	Test	—	1,722	1,721	874	0.51	1,721	847	0.49	1,721	843	0.49
Binary Classification	Train	Credible	41,671	39,824	23,170	0.58	34,040	20,379	0.60	39,445	21,853	0.55
Binary Classification	Train	Non-Cred.	9,353	8,272	5,734	0.69	5,496	4,034	0.73	7,831	5,040	0.64
Binary Classification	Val	Credible	3,117	2,960	1,669	0.56	2,503	1,446	0.58	2,940	1,566	0.53
Binary Classification	Val	Non-Cred.	3,117	2,717	1,929	0.71	1,784	1,321	0.74	2,619	1,694	0.65
Binary Classification	Test	Credible	3,117	2,967	1,739	0.59	2,555	1,543	0.60	2,947	1,629	0.55
Binary Classification	Test	Non-Cred.	3,117	2,713	1,920	0.71	1,782	1,328	0.75	2,620	1,704	0.65

Resources

• Repository
• Paper
• Common Crawl is our primary data source, supplemented with web scraping and multiple datasets for credibility signals:
  • DQR for credibility scores for supervised learning, and
  • Yasin et al.'s phishing domains,
  • Potpelwar et al.'s malware domains, and
  • Aung et al.'s legitimate domains, for semi-supervised learning.

Uses

This dataset is intended as a data source for research efforts against misinformation online. Specifically, as the first large-scale, text-attributed webgraph that is also dynamic, CrediBench stands as an ideal data source for efforts to develop methods for unreliable domain detection based on spatio-temporal cues.

Out-of-Scope Use

This dataset is not intended for LLM training. Designed for the goal of misinformation detection at the domain level and web scale, this dataset contains numerous domains and content pages that contain innapropriate content which may be harmful if used for training conversational AI, or other types of generative AI outside the scope of our task.

Data Collection and Processing

The process of collection, processing and use is detailed in our team's paper. We collect data through our proposed CrediBench pipeline (available at our repository), which builds a month's worth of data by pulling from Common Crawl, builds the graph from it and processes it to discard isolated and low-degree nodes. Each edge has a timestamp, given as the date of the first day of week of the crawl, in format YYYYMMDD.

Acknowledgements

• Curated by a team of collaborators from the Complex Data Lab @ Mila - Quebec AI Institute, the University of Oxford, McGill University, Concordia University, UC Berkeley, University of Montreal, and AITHYRA.
• Funding: This research was supported by the Engineering and Physical Sciences Research Council (EPSRC) and the AI Security Institute (AISI) grant: Towards Trustworthy AI Agents for Information Veracity and the EPSRC Turing AI World-Leading Research Fellowship No. EP/X040062/1 and EPSRC AI Hub No. EP/Y028872/1. This research was also enabled in part by compute resources provided by Mila (mila.quebec) and Compute Canada.
• License: CC-BY-4.0 (as retributed from Common Crawl).

Citation

BibTeX:

@article{kondrupsabry2025credibench,
  title={{CrediBench: Building Web-Scale Network Datasets for Information Integrity}},
  author={Kondrup, Emma and Sabry, Sebastian and Abdallah, Hussein and Yang, Zachary and Zhou, James and Pelrine, Kellin and Godbout, Jean-Fran{\c{c}}ois and Bronstein, Michael and Rabbany, Reihaneh and Huang, Shenyang},
  journal={arXiv preprint arXiv:2509.23340},
  year={2025},
  note={New Perspectives in Graph Machine Learning Workshop @ NeurIPS 2025},
  url={https://arxiv.org/abs/2509.23340}
}

APA:

Kondrup, E., Sabry, S., Abdallah, H., Yang, Z., Zhou, J., Pelrine, K., Godbout, J.-F., Bronstein, M., Rabbany, R., & Huang, S. (2025).
CrediBench: Building Web-Scale Network Datasets for Information Integrity.
New Perspectives in Graph Machine Learning Workshop @ NeurIPS 2025. arXiv:2509.23340. https://arxiv.org/pdf/2509.23340

Dataset Card Authors / Contact

For any questions on the dataset, please contact Emma Kondrup, Sebastian Sabry, or Shenyang (Andy) Huang.