README.md

---
license: other
license_name: us-government-public-domain
license_link: https://www.usa.gov/government-works
language:
  - en
size_categories:
  - 1M<n<10M
task_categories:
  - tabular-classification
pretty_name: "OPM FedScope Federal Employment — February 2026"
tags:
  - government
  - federal-workforce
  - foia
  - sampling-frame
  - opm
  - fedscope
configs:
  - config_name: default
    data_files:
      - split: train
        path: "employment_202602_part1.parquet"
---

# OPM FedScope Federal Employment — February 2026

## Why this dataset exists

We are building **FOIA requests** for federal agencies. To do that responsibly we need to
identify and sample the **custodians of records** likely to hold the documents we want —
i.e. the actual federal employees whose mailboxes, drives, and chat logs are responsive.

A naive list of senior officials over-samples the small visible top of an agency and
misses the bulk of work, which sits in the GS‑13/14/15 mass and the technical career
series. To get a defensible sample we want **stratification weights anchored in real
employment data**, so the FOIA targets reflect the actual distribution of staff across
sub-agencies, occupational series, pay plans, and locations.

This dataset is that anchor: OPM's record-level (PII-redacted) federal civilian
employment snapshot for **February 28, 2026** — every row is one employee.

> **Scope:** This is the full Feb‑2026 federal civilian workforce snapshot
> (**2,028,138 rows**, ~1.5 GB raw, compressed to 54 MB Parquet). The `_1_` in the
> source filename (`employment_202602_1_2026-05-04.txt`) is a legacy artifact from
> when OPM split the release into three parts; the current data.opm.gov download
> ships it as a single file. Source:
> [data.opm.gov/explore-data/data/data-downloads](https://data.opm.gov/explore-data/data/data-downloads).

## What's in here

- `employment_202602_part1.parquet` — ZSTD-compressed Parquet, **2,028,138 rows × 61 columns**, all columns typed as VARCHAR.
- The original pipe-delimited source from OPM (`employment_202602_1_2026-05-04.txt`) is not uploaded — it's reproducible from OPM and 28× larger.

### Schema (61 columns, all from OPM's published dictionary)

Identifiers / org:
`agency`, `agency_code`, `agency_subelement`, `agency_subelement_code`,
`cfo_act_agency_indicator`, `personnel_office_identifier_code`

Position:
`occupational_category` (P/A/T/C/B), `occupational_category_code`,
`occupational_group`, `occupational_group_code`,
`occupational_series`, `occupational_series_code`,
`pay_plan`, `pay_plan_code`, `grade`, `step_or_rate_type`, `step_or_rate_type_code`,
`position_occupied`, `position_occupied_code`,
`supervisory_status`, `supervisory_status_code`,
`appointment_type`, `appointment_type_code`,
`tenure`, `tenure_code`, `work_schedule`, `work_schedule_code`,
`flsa_category`, `flsa_category_code`,
`bargaining_unit`, `bargaining_unit_code`, `bargaining_unit_status`,
`nsftp_indicator`, `stem_occupation`, `stem_occupation_type`

Person attributes (coarse, no PII):
`age_bracket`, `length_of_service_years`, `education_level`, `education_level_bracket`,
`education_level_code`, `veteran_indicator`

Compensation (often `REDACTED`):
`annualized_adjusted_basic_pay`, `pay_basis`, `pay_basis_code`

Location:
`duty_station_code`, `duty_station_country`, `duty_station_country_code`,
`duty_station_county`, `duty_station_county_code`,
`duty_station_state`, `duty_station_state_abbreviation`,
`duty_station_state_country_territory_code`,
`core_based_statistical_area`, `core_based_statistical_area_code`,
`consolidated_statistical_area`, `consolidated_statistical_area_code`,
`locality_pay_area`, `locality_pay_area_code`,
`service_computation_date_leave`

Snapshot key:
`snapshot_yyyymm` (always `202602` here), `count` (always `1`)

### What this dataset will **not** give you

- **Free-text position titles** (e.g. "Chief of Staff", "Senior Advisor for Policy") — OPM strips these. Closest proxy is `occupational_series` (job family).
- **Personally identifiable information** — no names, no employee IDs.
- **Sub-office / front-office breakdown** — granularity stops at `agency_subelement` (e.g. all of "Office of the Secretary of the Interior" is one `IN01` bucket regardless of which office an employee actually sits in).
- **Adjusted basic pay** for many records (~122K of 418K Feb‑2026 VA records are null per OPM's release notes; redactions also common elsewhere).

For the named-position layer (Secretary, Deputy Secretary, Assistant Secretaries,
Schedule C / SES non-career, etc.) supplement this with the
[Plum Book](https://www.govinfo.gov/app/collection/plumbook), the agency org chart, and
SES/SL/ST listings.

---

## Recipe: building a FOIA custodian sampling frame with SQL

All examples use **DuckDB** against the Parquet file. Install with `pip install duckdb`
(or `uv run --with duckdb python ...`).

```python
import duckdb
con = duckdb.connect()
con.execute("CREATE VIEW emp AS SELECT * FROM 'employment_202602_part1.parquet'")
```

If you prefer the CLI:
```bash
duckdb -c "SELECT count(*) FROM 'employment_202602_part1.parquet'"
```

### 1. Pick the agency you're FOIA'ing

```sql
-- Find every subelement under DOI
SELECT DISTINCT agency_subelement, agency_subelement_code
FROM emp
WHERE agency_code = 'IN'
ORDER BY agency_subelement;
```

For DOI you'll see 14 subelements: `IN01` Office of the Secretary, `IN05` BLM, `IN06`
Indian Affairs, `IN07` Reclamation, `IN08` USGS, `IN10` NPS, `IN15` FWS, `IN21`
Solicitor, `IN22` OSMRE, `IN24` OIG, `IN26` BSEE, `IN27` BOEM, `IN28` BIE, `IN29` BTFA.

### 2. Get the headcount-per-subelement denominator (top of the stratification tree)

```sql
SELECT agency_subelement_code,
       agency_subelement,
       count(*) AS employees,
       round(100.0 * count(*) / sum(count(*)) OVER (), 2) AS pct_of_agency
FROM emp
WHERE agency_code = 'IN'
GROUP BY agency_subelement_code, agency_subelement
ORDER BY employees DESC;
```

Use the `pct_of_agency` column directly as your sub-agency sampling weight.

### 3. Within a subelement, stratify by tier and job family

A reasonable **Tier** proxy from `pay_plan_code` and `grade`:

| pay_plan_code | tier              |
|---|---|
| `EX`            | Presidentially appointed (PAS) |
| `ES`            | SES |
| `SL`, `ST`      | Senior Level / Senior Scientific |
| `GS` (grade ≥ 14), `GL` (≥ 14) | Senior career |
| `GS` (grade 12‑13)             | Mid career |
| `GS` (grade ≤ 11), `WG`, `WS`, `WL`, `WD` | Rank-and-file |
| else            | Other |

```sql
WITH tiered AS (
  SELECT *,
    CASE
      WHEN pay_plan_code = 'EX' THEN 'PAS'
      WHEN pay_plan_code = 'ES' THEN 'SES'
      WHEN pay_plan_code IN ('SL','ST') THEN 'SL_ST'
      WHEN pay_plan_code IN ('GS','GL') AND TRY_CAST(grade AS INT) >= 14 THEN 'Senior_career'
      WHEN pay_plan_code IN ('GS','GL') AND TRY_CAST(grade AS INT) BETWEEN 12 AND 13 THEN 'Mid_career'
      WHEN pay_plan_code IN ('GS','GL') AND TRY_CAST(grade AS INT) <= 11 THEN 'Rank_and_file'
      WHEN pay_plan_code IN ('WG','WS','WL','WD') THEN 'Rank_and_file'
      ELSE 'Other'
    END AS tier
  FROM emp
  WHERE agency_subelement_code = 'IN01'      -- swap to whatever you're FOIA'ing
)
SELECT tier,
       occupational_series_code,
       occupational_series,
       count(*) AS employees,
       round(100.0 * count(*) / sum(count(*)) OVER (PARTITION BY tier), 2) AS pct_within_tier
FROM tiered
GROUP BY tier, occupational_series_code, occupational_series
ORDER BY tier, employees DESC;
```

### 4. Build your sampling frame as a single tidy table

```sql
COPY (
  SELECT
    agency_code,
    agency_subelement_code   AS office_code,
    agency_subelement        AS office,
    pay_plan_code,
    grade,
    occupational_category_code,
    occupational_series_code AS series_code,
    occupational_series      AS series,
    duty_station_state_abbreviation AS state,
    count(*)                 AS employees
  FROM emp
  WHERE agency_code = 'IN'
  GROUP BY ALL
) TO 'doi_sampling_frame.parquet' (FORMAT PARQUET);
```

Now you can allocate FOIA-custodian sample slots proportional to `employees` (or any
weighted scheme — e.g. over-sample SES tiers, then use this frame to generate quotas for
the GS‑13/14/15 mass).

### 5. Sanity-check the frame against what OPM publishes

```sql
SELECT count(*) AS rows, sum(count(*)) OVER () AS employees
FROM emp WHERE agency_code = 'IN';
```

`employees` should match OPM's published DOI total for Feb 2026.

---

## Recreating this file from OPM

1. Visit <https://data.opm.gov/explore-data/data/data-downloads>.
2. Click **DOWNLOAD** under "Federal Employment Raw Data (February 2026)". Note: this is
   a Blazor Server app that streams the file over WebSocket; large transfers may
   disconnect under headless automation. Real browsers work.
3. The download is a `.txt` file with `|` delimiters and a header row, named
   `employment_202602_<part>_<download-date>.txt`.

To regenerate the Parquet:
```bash
uv run --with duckdb python -c "
import duckdb
duckdb.sql('''
  COPY (SELECT * FROM read_csv(\"employment_202602_1_*.txt\",
                               delim=\"|\", header=true, all_varchar=true))
  TO \"employment_202602_part1.parquet\" (FORMAT PARQUET, COMPRESSION ZSTD)
''')
"
```

## Provenance & license

- **Source:** U.S. Office of Personnel Management, Enterprise Human Resources
  Integration (EHRI) Status snapshot, published via
  [data.opm.gov](https://data.opm.gov/).
- **Coverage:** Federal civilian workforce snapshot as of **2026-02-28**, published
  **2026-03-30**, version 1.
- **PII:** OPM redacts personal identifiers and many compensation values. No employee
  names, no employee IDs.
- **License:** U.S. federal government work — public domain in the U.S. under
  [17 U.S.C. § 105](https://www.usa.gov/government-works). Not endorsed by or affiliated with OPM.

## Citation

```
U.S. Office of Personnel Management (2026). Federal Employment Raw Data — February 2026.
Enterprise Human Resources Integration (EHRI) Status dataset.
https://data.opm.gov/explore-data/data/data-downloads
```