Flash floods develop within minutes to hours, occur mainly in small to medium-sized watersheds, and are documented through scattered sources such as news reports, social media, stream gauges, operational storm events databases, and post-flood field surveys. Such fragmentation historically limits flash flood research advances and model evaluation or intercomparisons. We present <strong>TORRENT</strong> (the <strong>T</strong>raceable, <strong>O</strong>bservation-constrained, <strong>R</strong>apid-<strong>R</strong>esponse, <strong>E</strong>pisode–gauge–watershed <strong>N</strong>etwork of <strong>T</strong>estbeds), a benchmark-ready community dataset of flash-flood hydrological response in the contiguous United States (CONUS). Each testbed links one reported NOAA/NCEI flash-flood episode, one U.S. Geological Survey (USGS) stream gauge, and the gauge’s contributing watershed, combining the observed discharge hydrograph—the primary benchmark target—with watershed boundaries, temporally screened in-basin high-water marks, and quality-control metadata; full packages additionally provide Multi-Radar Multi-Sensor (MRMS) precipitation forcing and auxiliary gridded products. TORRENT follows a three-level structure distinguishing what was reported, what observations are available, and what can be evaluated reproducibly: a national inventory of 42,466 reported episodes for 1996–2025; observation-availability documentation for 5,424 recent CONUS episodes during 2021–2025, including episodes without usable external observations; and 806 episode–gauge–watershed evaluation packages representing 160 episodes and 208 gauges. Each Level-3 package has a retrievable watershed, at least one in-basin high-water mark meeting the stated temporal-screening criterion, and a detected discharge response. Flash-flood-response screening, deduplication, in-basin NCEI corroboration, station-relative severity screening, and post-selection curation yield a conservative primary benchmark of 21 testbeds representing 14 episodes and 21 gauges. This primary set is intended as a high-confidence reference for formal rainfall–runoff model intercomparison rather than as a statistically representative sample of CONUS flash floods; the broader Level-3 pool supports exploratory analysis and robustness testing. Primary testbeds exhibit rapid, pronounced responses, with estimated event-peak return periods ranging from about 2 to more than 100 years. The automated construction and screening stages are released as deterministic, re-runnable code, together with documented curation decisions, sensitivity tests, and quality-control diagnostics. TORRENT is openly available at <a href="https://doi.org/10.5281/zenodo.23359069" target="_blank" rel="noopener">https://doi.org/10.5281/zenodo.23359069</a> (Yan et al., 2026).