A clean, deduplicated, full-field corpus of electronically filed IRS Form 990 and 990-EZ returns, parsed from the official IRS e-file XML into 15 analysis-ready parquet tables — every schedule, every field, 6.2M filings, 160.4M rows, tax years 2010–2026.

Query it from anywhere in one line of DuckDB — no download, no signup, no server:

INSTALL httpfs; LOAD httpfs;
SELECT ein, F9_00_ORG_NAME_L1, F9_09_EXP_TOT_TOT
FROM read_parquet('hf://datasets/mauricedalton/irs-990/main/tax_year=2023/*.parquet')
WHERE ein = '222529464';

Why this dataset

A parsed, query-ready view of the full Form 990/990-EZ e-file record: every schedule and every field, deduplicated to one row per filing with an explicit audited rule (see Provenance), covering tax years 2010–2026 — including pre-2019 filings recovered from NBER archives — and published as columnar parquet you can query remotely with no download.

Tables

Table	Contents	Grain	Rows
`main`	Core 990/990-EZ return (693 cols: parts I–XII)	1 row / filing	6,207,891
`person`	Officers, directors, trustees, key employees (Part VII)	repeating	61,695,135
`scheda`	Schedule A — public charity status	repeating	5,310,504
`schedd`	Schedule D — supplemental financials	repeating	16,768,776
`schedf`	Schedule F — foreign activities	repeating	1,864,896
`schedh`	Schedule H — hospitals	repeating	655,410
`schedi`	Schedule I — domestic grants	repeating	11,464,694
`schedj`	Schedule J — executive compensation	repeating	4,915,270
`schedk`	Schedule K — tax-exempt bonds	repeating	271,135
`schedl`	Schedule L — interested-party transactions	repeating	1,049,788
`schedm`	Schedule M — noncash contributions	repeating	1,369,108
`schedn`	Schedule N — liquidation/dissolution	repeating	200,282
`schedo`	Schedule O — supplemental free-text	repeating	33,531,646
`schedr`	Schedule R — related organizations	repeating	14,791,482
`facilities`	Hospital facilities (Sch. H Part V)	repeating	323,335

Every table is hive-partitioned by tax_year and sorted by ein within each partition, so EIN- and year-filtered remote queries touch only the row groups they need. Join any schedule to main on object_id (one e-filed return) or analyze by ein + tax_year.

A 16th config, raw_xml, carries the original e-file XML behind these parsed tables — for re-parsing, auditing, or working with returns as filed.

Usage

DuckDB (SQL or Python) — streams only the bytes your query needs:

import duckdb
con = duckdb.connect()
con.execute("INSTALL httpfs; LOAD httpfs;")
df = con.sql("""
    SELECT tax_year, count(*) AS filings, sum(SH_01_FA_EXP_NET_COM_BEN) AS fin_assistance_net
    FROM read_parquet('hf://datasets/mauricedalton/irs-990/schedh/tax_year=*/part-*.parquet')
    GROUP BY 1 ORDER BY 1
""").df()

🤗 Datasets:

from datasets import load_dataset
ds = load_dataset("mauricedalton/irs-990", "schedj", split="train")

pandas / Polars (single partition):

import pandas as pd
df = pd.read_parquet("hf://datasets/mauricedalton/irs-990/main/tax_year=2023/part-0.parquet")

Archival raw XML

The raw_xml config ships the original IRS e-file XML behind the 15 parsed tables — pull it to re-extract fields this corpus doesn't surface, audit the parsing, or work with returns exactly as filed:

INSTALL httpfs; LOAD httpfs;
SELECT object_id, source, archive_year, raw_xml
FROM read_parquet('hf://datasets/mauricedalton/irs-990/raw_xml/year=2023/*.parquet')
WHERE ein = '222529464';

It is deliberately different from the 15 analytical tables in three ways:

	Analytical tables	`raw_xml`
Grain	1 row / filing (deduped)	1 row / retained copy — `(object_id, source, archive_year)`, not deduped
Partition	`tax_year=`	`year=` (tax year parsed from the XML)
Payload	parsed columns	one `raw_xml` BLOB per row + filing keys
Size	9.5 GB	24 GB — 7.10M rows, 614 files

Columns: object_id, ein, return_type_cd, tax_year, tax_period_end, source (irs_direct / nber), archive_year, schema_version_detected, ingest_timestamp, raw_xml (BLOB), year. Each file is sorted by object_id. Because raw_xml keeps every retained copy of a filing, apply the dedup rule yourself if you need one row per filing (latest archive_year wins; irs_direct beats nber on ties). The XML blobs are large — query with DuckDB (it streams only the row groups you touch) rather than the dataset viewer.

Provenance & method

Sources: NBER archival mirrors of the IRS e-file releases (covering the pre-2019 files irs.gov no longer hosts) + IRS-direct downloads for 2019–2026. Every source archive is checksummed and retained.
Parsing: field extraction driven by the NODC concordance plus supplemental mappings, applied uniformly across XML schema versions (including post-2018 versions).
Dedup: one row per filing: latest archive year wins; on ties, IRS-direct beats NBER; content-hash comparison within object_id. A per-object_id provenance ledger is kept upstream.
Validation: multi-gate integrity check (row counts, grain uniqueness, EIN/year coverage) passes on the published build; parquet export row counts are verified against the build manifest (export_manifest.json, including the source-build git commit).

Scope & limitations

Form 990 and 990-EZ only (3.97M / 2.24M filings). No 990-PF (private foundations), 990-T, or 990-N postcards.
E-filed returns only — paper-only filers appear from their first e-file year.
tax_year 2025–2026 are partial (filing-season lag).
4 rows carry a pre-2010 tax_year (tax_year=2000: 1 row, tax_year=2001: 3 rows in main) — TaxYr typos in the source XML of four 2020-processed filings, kept verbatim for archival fidelity. Filter tax_year >= 2010 for tax-year analysis.
Column names follow the concordance-style convention: F9_<part>_<field> on main (e.g. F9_09_EXP_TOT_TOT), SH_…/SJ_…/etc. on schedules; every table also carries ein, object_id, return_type_cd, tax_year. See the per-table parquet schema for the full list.

Recent extraction corrections

Schedule H Part IV (management companies & joint ventures) is now extracted — the SH_04_* columns are populated on rows with part = 'Part_IV_Mgmt_Co_JV' (previously always null).
Schedule R legacy ownership — SR_03/SR_04_*_PCT_OWN and *_EOY_ASSETS are now populated for pre-2013 filings (2009–2012 vintages named these elements differently and were previously dropped).
main Part V compliance counts — F9_05_NUM_EMPL (W-3 employee count) and F9_05_NUM_FORM_1096_BOX_3 (1099/1096 count) are now populated for 2017+ filings (previously null past the concordance's 2016 vintage cutoff).
main modern-vintage fallbacks — 200+ F9_* columns that read null for 2017+ filings (the concordance's version lists end at 2016) are now populated via schema-version fallbacks.
main program-service values corrected — an xpath-precedence fix routes F9_03_PROG_{EXP,DESC,GRANT,CODE} from the right element; a handful of rows that previously published a different program's figure now carry the correct line-4a value.
Schedule K for pre-2013 filings — Parts I–V (SK_*) are now extracted for 2009–2012 bond filings (previously unextracted); advance/current-refunding indicators are now populated for 2011–2017 vintages.
Schedule A Part V — the Type III supporting-organization distribution columns (SA_05_*) are now populated for 2018+ filings.
Schedule H Part V Section B — 100+ hospital facility policy columns (SH_05_*/SH_99_*) are now mapped across both eras.
Schedule R / Schedule I legacy fields — pre-2013 addresses, domicile, transaction, and recipient columns recovered.
Known upstream limitation (not an extraction gap): Schedule R *_PCT_OWN fill drops from ~12–13% (2014–2018) to ~8% (2019+) in the IRS-published XML itself — the element is genuinely absent, not dropped by the parser (verified against raw XML). Post-2018 ownership-percentage missingness is missing-not-at-random at the source; model the selection or restrict to 2013–2018.

License & citation

Underlying 990 filings are US public records; this compilation is released CC0 1.0. If you use it in research, please cite:

@dataset{irs990corpus2026,
  title  = {IRS Form 990 e-file corpus (2010--2026)},
  author = {Maurice Dalton},
  year   = {2026},
  url    = {https://huggingface.co/datasets/mauricedalton/irs-990}
}

Downloads last month: 34

Total file size:

35.5 GB