GAO-Appropriations / README.md
Kmisener's picture
Upload 2 files
11cf914 verified
|
Raw History Blame Contribute Delete
10.8 kB
metadata
license: other
license_name: us-government-work
license_link: https://www.gao.gov/about/website-policies
language:
  - en
pretty_name: GAO Appropriations Law (Fiscal Law) Decisions
size_categories:
  - 10K<n<100K
task_categories:
  - text-retrieval
  - text-classification
  - question-answering
tags:
  - legal
  - government
  - appropriations
  - fiscal-law
  - us-federal
configs:
  - config_name: default
    data_files: gao_appropriations_corpus.jsonl
dataset_info:
  config_name: default
  features:
    - name: url
      dtype: string
    - name: file_slug
      dtype: string
    - name: title
      dtype: string
    - name: b_numbers
      dtype: string
    - name: decision_date
      dtype: string
    - name: decision_year
      dtype: int64
    - name: released_date
      dtype: string
    - name: agency
      dtype: string
    - name: subject
      dtype: string
    - name: requester
      dtype: string
    - name: pillars
      sequence: string
    - name: doctrines
      sequence: string
    - name: statutes
      sequence: string
    - name: conclusion
      dtype: string
    - name: digest
      dtype: string
    - name: highlights
      dtype: string
    - name: text_source
      dtype: string
    - name: text_noise_pct
      dtype: float64
    - name: text_quality
      dtype: string
    - name: pdf_file
      dtype: string
    - name: text
      dtype: string
  splits:
    - name: train
      num_bytes: 551999106
      num_examples: 67246

GAO Appropriations Law (Fiscal Law) Decisions

67,246 decisions of the U.S. Comptroller General and the Government Accountability Office on the availability of appropriations — purpose, time and amount — spanning 1922 to 2026. 481.2M characters of full decision text, mean 7,156 characters per decision.

This is the complete published Appropriations Decision series as GAO's own catalogue defines it, not a sample.

Loading

from datasets import load_dataset

ds = load_dataset("Kmisener/GAO-Appropriations", split="train")
print(ds[0]["title"], ds[0]["decision_year"])

The schema is declared explicitly in this card's features block. That is deliberate and load-bearing: several columns (decision_date, conclusion, requester, agency, digest) are null for the first thousands of rows because the corpus is ordered oldest-first and those fields come from section headings that pre-1995 decisions do not use. Without an explicit schema, Arrow infers them as null from the first block and then fails on the first real value with Column(/decision_date) changed from null to string.

Coverage, and what is deliberately not here

Pass 1 enumerated the catalogue in 702 date slices, each verified against the item count GAO prints for it: 702 slices, 68,087 items, 0 short. Pass 2 fetched 68,132 of 68,138 indexed URLs, or 99.99%; the 6 missing are dead -0 slugs that returned nothing on two attempts. 886 decisions with under 500 characters of text were then dropped, leaving 67,246.

This covers one leaf of GAO's ctype_search taxonomy. GAO's legal output is a two-level tree, and Appropriations Decision is one branch of it:

branch items here
Appropriations Decision 68,086 yes, all of it
Antideficiency Act Report 23 no
Federal Agency Major Rule Report — Congressional Review Act 2,197 no
Other Decision — includes CRA legal opinions 503 no
Legal Other Written Product 2,297 no
Bid Protest Decision 33,292 separate dataset

Congressional Review Act material is not in this dataset. CRA compliance reports on major rules live under Federal Agency Major Rule Report, and — more easily missed — CRA legal opinions are filed under Other Decision, whose first result is "U.S. Department of Education—Applicability of the Congressional Review Act". Those opinions read like appropriations decisions and reason about whether something is a rule under CRA, but GAO does not classify them as appropriations decisions.

The corpus is broader than "fiscal law"

Appropriations Decision carries the whole historical Comptroller General decision series, not only fiscal-law questions. Measured by vocabulary over all records: about 36% use appropriations vocabulary at all, about 16% carry a strong bid-protest signal, and roughly half match neither — they are older decisions on claims, military and civilian pay, travel allowances and similar.

So topic filters, not the file itself, define a fiscal-law dataset:

# the practical fiscal-law subset: 21,560 decisions (32.1%)
fiscal = ds.filter(lambda r: len(r["doctrines"]) > 0)

# stricter -- a named doctrine AND a statutory citation: 8,032 (11.9%)
strict = ds.filter(lambda r: r["doctrines"] and r["statutes"])

statutes alone is non-empty for 18,043 records (26.8%).

Schema

field type populated notes
url string 100% canonical gao.gov product URL
file_slug string 100% on-disk stem, e.g. b-155290
title string 100%
b_numbers string 100% e.g. B-336373
subject string 100% the title, or its right-hand side
highlights string 100% GAO's own summary or published digest
text string 100% the complete decision, footnotes and signature included
text_source string 100% page (60,313) or pdf (6,933)
text_noise_pct float64 100% share of characters that look like OCR noise
decision_year int64 97.7% falls back to published/released date
released_date string 58.1%
doctrines list<string> 32.1% non-empty named rules used or merely cited
statutes list<string> 26.8% non-empty normalised, e.g. 31 U.S.C. § 1502(a)
pdf_file string 11.5% GAO hosts PDFs only for recent decisions
pillars list<string> 5.0% non-empty 3,392 records: purpose (2,516), time (609), amount (668)
digest string 3.4% from a DIGEST heading
agency string 1.6% from the title, before the em dash
requester string 1.5% congressional requests only (The Honorable …)
text_quality string 1.3% ocr_suspect where set (855 records)
conclusion string 0.9% GAO's own CONCLUSION section, verbatim
decision_date string 0.4% from a Date: line — not reliably a bare date, see below

Why the low rates are not a defect

The corpus' centre of gravity is far older than people expect: 1920s 2,214; 1930s 3,366; 1940s 3,320; 1950s 9,937; 1960s 12,420; 1970s 10,046; 1980s 15,095; 1990s 7,705 — and only 1,606 decisions after 2000.

agency, decision_date, digest and conclusion all depend on conventions that are recent: an <entity>—<question> title, a Date: line, a DIGEST or CONCLUSION heading. Pre-1995 decisions use none of them. Titles are instead B-228838 Sep 16 1987, or a bracketed abstract, or a plain sentence. Do not filter a historical subset by agency — it will silently drop almost everything. Use doctrines, statutes or a text match, which work across every era.

highlights is the field to read for historical decisions: it is 100% populated and carries GAO's own published digest.

Four things to know before trusting a field

  1. doctrines is a search index, not a set of holdings. A hit means the decision used the phrase or cited the statute, not that GAO decided on that ground. A single bona fide needs case will also mention the Antideficiency Act and the recording statute. Use it to narrow, then read.
  2. There is deliberately no outcome field. Appropriations decisions answer a legal question in prose, and the answer is routinely split ("AOC properly obligated its FY 2023 funds … AOC may not use its annual appropriation for X"). On the sibling protest corpus, a classifier reading the end of each decision agreed with GAO's own labels only 48% of the time. Read conclusion, which is GAO's words, and label it yourself against a hand-checked sample.
  3. The page, not the PDF, is the better text source — the reverse of what you would expect. GAO hosts full-report PDFs only for recent decisions; the old ones are scans with poor OCR ("COMP TflOLLE"R GENERAL"). The old pages are clean keyed text. 855 records are flagged ocr_suspect.
  4. decision_date is not reliably a bare date — do not parse it blindly. Of the 272 records where it is set, 199 (73.2%) parse as a date. The other 73 over-captured the lines following GAO's Date: field, producing values like "September 29, 1988 To: Judgment Group Manager, Claims/GGD-Kenneth Schutt Attention: Rufus Hill". Use decision_year (97.7%, an integer) for anything chronological, and treat decision_date as a text hint. The leading date is correct in every polluted value inspected, so a leading-date regex recovers it if you need the day.

Known limitations

  • 6 decisions in GAO's catalogue have no page and are absent; each was attempted twice.
  • 886 decisions with under 500 characters of text were dropped from this cut. They exist in GAO's catalogue.
  • GAO's printed item counts run 1–4 above what pagination returns, so the catalogue total drifts by a handful day to day as GAO publishes.
  • Text is GAO's HTML body only — no navigation, footer or related-links furniture ever entered it. GAO's protective-order notice is stripped; it is a protest artifact and is rare here.
  • decision_year is absent for 1,537 records that carry no usable date in any of the three date fields.

Provenance

Scraped from www.gao.gov between 2026-10-26 and 2026-11-05 directly from GAO's own faceted search (ctype_search:Appropriations Decision), one page snapshot and one PDF per decision retained, with every structured field re-derived offline from the saved HTML. Enumeration was date-sliced and every slice reconciled against the count GAO prints for it, because paging the catalogue flat leaks about 3% of results through unstable ordering.

Crawl pacing was deliberately conservative and GAO's published rate limits were respected.

License

GAO decisions are works of the United States Government and are not subject to copyright protection in the United States (17 U.S.C. § 105). This compilation adds no additional restrictions. Attribution to the U.S. Government Accountability Office is appropriate but not legally required.

Citation

@misc{gao_appropriations_decisions,
  title  = {GAO Appropriations Law (Fiscal Law) Decisions},
  author = {U.S. Government Accountability Office},
  note   = {Compiled corpus of 67,246 decisions, 1920--2026},
  year   = {2026}
}