mauricedalton/cms-nppes

CMS NPPES — historical annual snapshots

Historical CMS National Plan and Provider Enumeration System (NPPES) files ingested from the NBER NPPES mirror (https://data.nber.org/npi) and staged as annual, analysis-ready Parquet (zstd-compressed). The pipeline crawls the NBER directory, downloads the source files, converts and unifies them, then validates the National Provider Identifier (NPI) via Luhn check, uniqueness, and a cross-check against the raw CMS extract before promoting to data/final/. Four products are published per year under data/final//_.parquet: 'core' (the analytic workhorse, one row per NPI), 'npi' (the full NPPES extract), 'othpid' (the unified other-provider-identifier crosswalk family), and 'statelic' (state license records). Products are published as parquet only; a bundled converter (python -m nppes to-stata) produces labeled Stata .dta files on demand. Provenance caveat: for vintages after 2017, source data on the NBER mirror is distributed as Stata .dta rather than the original CMS dissemination CSVs, so those years inherit NBER's variable typing/labeling; the validation stage's raw-CMS cross-check guards against drift introduced by that conversion.

Subsets (configs) and snapshots (splits)

Each product is a config; each annual snapshot is a split named y<YEAR> (the snapshot year is otherwise carried only in the filename).

config	years	snapshots
core (default)	2007–2026	18
npi	2007–2026	18
othpid	2007–2026	18
statelic	2007–2026	17

Usage

from datasets import load_dataset

# one product, one annual snapshot
core_2017 = load_dataset("mauricedalton/cms-nppes", "core", split="y2017")

# a whole product (all years) — iterate the snapshots
core_all = load_dataset("mauricedalton/cms-nppes", "core")

Read a single file directly (no auth — public dataset), e.g. with DuckDB:

INSTALL httpfs; LOAD httpfs;
SELECT count(*) FROM read_parquet('hf://datasets/mauricedalton/cms-nppes/core/core_2017.parquet');

or pandas / polars over hf://:

import pandas as pd
pd.read_parquet("hf://datasets/mauricedalton/cms-nppes/core/core_2017.parquet")

Provenance

Ingested from the NBER NPPES mirror and staged as annual, analysis-ready Parquet (zstd). NPIs are Luhn-validated and reconciled against the raw CMS dissemination archive.

A raw CMS dissemination product is used internally as a validation cross-check and is not published here.

License

cc0-1.0 — the underlying NPPES data is a U.S. Government work in the public domain.