Collective Vigilance Intelligence Dataset for Risk Analysis and Governance Auditing

Collective Vigilance Intelligence is a research dataset containing 12,000 event-level units for analyzing complex risk information. It was created by MENG, WEI and integrates three layers: 7,200 real public-source samples, 3,000 derived synthetic samples, and 1,800 multimodal composite samples. The dataset was last updated in April 2026.

Use Cases

• Train models for cross-lingual robustness evaluation using event logs and cross-lingual retrieval settings from public-source data layers like GDELT 2.0 and MIRACL.
• Test resilience of verification systems against semantic perturbation and relation fabrication conditions in the derived synthetic data layer.
• Analyze governance measures such as the Governance Compatibility Index, traceability, and explanation consistency recorded in the audit layer.
• Evaluate multimodal consistency analysis by combining event-relation structures from MAVEN-ERE with multimodal crisis samples from CrisisMMD.
• Study bias sensitivity and verification outcomes using audit layer records of human corrections and final decision rationales.

Strengths

• Contains 12,000 structured event-level units with a predefined 70/15/15 train/validation/test split.
• Integrates data from four established public sources (GDELT 2.0, CrisisMMD, MAVEN-ERE, MIRACL) covering event logs, multimodal samples, and cross-lingual settings.
• Includes a dedicated audit layer with governance measures like traceability and bias sensitivity for accountability analysis.

Limitations

• Contains 3,000 rule-generated synthetic samples, which may not fully capture the complexity of real-world risk events.
• Specific geographic coverage and precise time range of the public-source data are not detailed, limiting reproducibility for spatiotemporal studies.
• Access to selected processed materials requires a request and compliance review, potentially hindering immediate use.

Provenance

Source

Integrated from GDELT 2.0, CrisisMMD, MAVEN-ERE, and MIRACL public-source datasets.

Collection Method

Combines real public-source data, rule-generated derived synthetic data, and governance-audit records.

Time Range

null

Freshness

null

Geography

null