Nason/dol-visas-database

DOL Visas Database (H-1B LCA + PERM)

Every H-1B/H-1B1/E-3 Labor Condition Application and PERM permanent labor certification application disclosed by the DOL Office of Foreign Labor Certification, FY2015 to present, as a single queryable DuckDB database. 8,812,639 rows across 2 tables.

Table	Description	Row Count	Column Count	Date Range
lca	H-1B/H-1B1/E-3 Labor Condition Applications, one row per application per disclosure file, FY2015-present	7,479,697	110	FY2015 to FY2026
perm	PERM permanent labor certification applications, one row per application per disclosure file, FY2015-present	1,332,942	73	FY2015 to FY2026

Query it remotely

INSTALL httpfs; LOAD httpfs;
ATTACH 'https://huggingface.co/datasets/Nason/dol-visas-database/resolve/main/dol_visas.duckdb' AS visas (READ_ONLY);

SELECT employer_name, COUNT(*) AS certified_lcas
FROM visas.lca
WHERE fiscal_year = 2025 AND case_status = 'CERTIFIED' AND is_latest
GROUP BY 1 ORDER BY 2 DESC LIMIT 10;

Or with the datapond packages: pip install datapond / pak::pak("datapond-db/datapond-r").

Build pipeline, per-era column crosswalks, and full documentation: https://github.com/ian-nason/dol-visas-database

Changelog

2026-09-14 — Data refresh (FY2026 Q3)

• Rebuilt from the OFLC FY2026 Q3 disclosure files (determinations through 2026-06-30, posted August 2026), replacing the cumulative Q2 files. Total rows 8,533,076 → 8,812,639: lca 7,479,697 (FY2026 to date 437,496, was 259,200), perm 1,332,942 (FY2026 to date 112,550).
• Fix: date columns were 100% NULL in the July release. DuckDB's read_xlsx(all_varchar=true) returns Excel date serials as text (45931), which the DATE cast silently rejected, so decision_date, received_date, original_cert_date and orig_file_date were empty in both tables (and begin_date/end_date partly). The cast now converts serials, and validation fails if decision_date/received_date are not populated.
• DOL made no layout changes: the Q3 LCA (98 columns) and PERM (137 columns) headers are identical to Q2, so the crosswalks are unchanged.
• Builder now fails loudly if a source file carries a column that is neither in the crosswalk nor in the new known_unmapped.json (the 464 columns we intentionally drop, e.g. FY2019 secondary-worksite groups 2-10); previously unknown columns were dropped silently. Validation now also asserts every fiscal year FY2015-FY2026 is present and is_latest is unique per case.
• DuckDB memory limit lowered to 5 GB (was 8 GB) with 4 threads so the build fits the 10 GB build VM; publish_to_hf.py --card-only re-uploads the dataset card without re-uploading the file.

2026-07-07 — Initial release

• lca: H-1B/H-1B1/E-3 Labor Condition Applications, FY2015 through FY2026-to-date (~7.25M rows, 106 harmonized columns from 4 source-layout eras).
• perm: PERM labor certification applications, FY2015 through FY2026-to-date (~1.28M rows, 69 curated analytical columns from 3 form eras; remaining source columns documented as drops in the crosswalk).
• All disclosure rows kept; is_latest flags the most recent appearance of each case number per program.
• Categorical domains normalized across eras (case_status casing, wage levels, Y/N flags, literal 'NULL' strings).
• Known upstream limits: FY2024+ PERM form drops citizenship/admission-class and foreign-worker education fields; FY2019 LCA secondary worksite groups (2-10) not retained; H-2A/H-2B/CW-1/prevailing-wage files out of scope.