An open, multi-state, depth-resolved inventory of per- and polyfluoroalkyl substances (PFAS) in soil, harvested from agency investigation reports with vision-language models

Depth-resolved PFAS measurements in soil, harvested from public U.S. agency site-investigation reports with a vision-language-model extraction pipeline. Companion dataset to a manuscript at Environmental Science & Technology. Contents: 30,972 measurement records (30,748 canonical); 6,779 samples; 5,812 borings; 141 public sources; 41 U.S. states; 72 distinct PFAS; 23,060 records with a sanity-passed depth interval. Ten QA gates; independent audit (n=160): 94.4% value presence, 100% depth-interval agreement, 1.7% hallucination. Soil only (no groundwater/surface-water/sediment/private wells). Includes the 691-site AFFF/defense census denominator. Data CC-BY-4.0; code MIT.