The candidate-level evaluation frame for the hand-labeled 1,502-organization
benchmark: one row per (source organization x BMF candidate), joined with the
ground-truth labels. Group rows by uei to see a source org's competing
candidates; is_top_candidate == 1 marks the algorithm's chosen pick and
is_gt_ein == 1 marks the row whose EIN is the ground-truth correct match.
Source
SAM / USASpending nonprofit registrations matched to the IRS Business
Master File (unified active + inactive, 2026-01 vintage). See
vignette("training-dataset", package = "npmatch").
Details
Alongside similarity features it carries the full review evidence: normalized
name variants, IDF-annotated tokens, geographic-agreement flags, the veto
trail, imported BMF context (BMF_*) and SAM context (SAM_*), and the
ground-truth columns (gt_*). Every column is described in
vignette("training-dataset", package = "npmatch") and the field-level
vignette("candidate-evaluation-frame", package = "npmatch").