RSKGD-FL-2014-2018: A Five-Year Florida Temporal Road Safety Knowledge Graph Dataset and Benchmark

RSKGD-FL-2014-2018 is a five-year temporal road safety knowledge graph dataset and benchmark covering Florida crash events from 2014 through 2018. The release contains 2,077,333 canonical crash events assembled from frozen snapshots of the Florida Department of Transportation Crashes_All service and annual crash-independent roadway universes derived from U.S. Census Bureau TIGER/Line All Lines (Edges) vintages. The deposited package provides documented Bronze, Silver, and Gold data layers; annual roadway, node, edge, and graph tables; crash–road linkage outputs; provenance records; schemas; data dictionaries; manifests; quality-control reports; and prospective temporal benchmark definitions. It includes an as-of roadway universe for prospective evaluation and a retrospective multi-year universe for descriptive and lineage analyses. Road units with zero linked crashes are retained. Crash-to-road assignment is algorithmic. The package distinguishes selected, candidate, and unmatched relations, and selected links should not be interpreted as adjudicated ground truth. The release does not provide a complete exposure denominator for the full roadway universe and should not be used to infer exposure-normalized crash risk without additional data. The dataset supports research on temporal road safety knowledge graphs, rare-event prediction, graph and tabular learning, crash–road linkage, change of support, temporal benchmark design, provenance, and leakage-controlled prospective evaluation. It is intended for reproducible prediction and benchmark research rather than causal inference. This is an independent five-year release. It is related to, but is not a new version of, the one-year, two-year, ContextPlus, or New York City RSKGD datasets. Licensing is component-specific; users should consult LICENSES_AND_TERMS.md in the deposited archive.