Foreign Relations of the United States (FRUS)

The Foreign Relations of the United States series is the official documentary record of US foreign policy, compiled and annotated by the US Department of State's Office of the Historian (OH). It runs from 1861 to the 1980s and is made up of telegrams, memoranda of conversation, meeting minutes, letters, intelligence estimates and editorial notes. This dataset turns OH's TEI XML master files into document-level Parquet tables. (TEI, the Text Encoding Initiative, is the XML vocabulary used for scholarly editions.) It keeps the volume and chapter structure, dates, source notes and footnotes, and includes each document's original TEI.

• Upstream commit: 6170e4b32c12 (2026-10-14T00:18:24Z), latest release v1.1.4
• Built: 2026-10-29T16:48:42Z (build.py parser v1)
• Volumes: 744 (553 with documents; status — published: 551, planned: 81, being-cleared: 54, being-researched: 45, being-digitized: 10, partially-published: 3)
• Documents: 314,616
• Glossary entries: 128,708; front/back-matter sections: 1,756
• Parquet size: 1.00 GB

Source

• Upstream: https://github.com/HistoryAtState/frus. volumes/<volume_id>.xml holds one TEI P5 file per volume (744 files, 3.3 GB).
• Online edition: https://history.state.gov/historicaldocuments
• Snapshot: master commit and release shown above, retrieved 2026-10-28. The commit SHA, and the git blob SHA of every volume file, are also recorded in sync_state.json.

How this differs from other copies

HistoryLab/foiarchive also contains FRUS. It is gated, licensed CC BY-NC, flattens the documents into one corpus and is a 2024 snapshot. This dataset is CC0, built straight from OH's TEI masters, keeps volume, chapter and footnote structure, and is kept in sync with upstream.

Configs

documents (default): one row per historical document

One row is one TEI <div type="document">: one numbered document in a FRUS volume, or an editorial note. Nested sub-documents never occur in the current data (parent_doc_id is always null). Attachments, enclosures and "Tab A"-style sub-parts are part of their parent document's text.

column	description
volume_id	Volume identifier, e.g. frus1969-76v24 (the upstream file name)
doc_id	Canonical @xml:id of the document within its volume, e.g. d176. Usually d<n>. A few volumes use other ids (bl_d1, d1004a-1, appA)
url	https://history.state.gov/historicaldocuments/<volume_id>/<doc_id>
doc_number	Printed document number (@n), as a string ("176", "62a")
subtype	historical-document (306,148 rows) or editorial-note (8,468 rows)
ordinal	0-based position of the document within the volume
part	body, or front/back for a few documents printed in front or back matter
volume_title, subseries, administrations, volume_status	Copied from the volume header. subseries is the @n value, e.g. 1969-76. administrations are OH's coverage keywords, e.g. nixon-ford
path	Nesting of enclosing divs, outermost first: list of {id, type, subtype, n, title}. type is compilation, chapter, subchapter and so on. Depth runs from 0 to 7
chapter_title	Title of the innermost enclosing div that has one
parent_doc_id, nested_doc_ids	Included for safety; both are empty in the current snapshot
title	Document heading as plain text, with footnote markers removed, e.g. 176. Intelligence Note From … to Secretary of State Rogers
date_min, date_max	Normalised date range (date32), taken from OH's frus:doc-dateTime-min/max attributes. Every document has them
datetime_min_raw, datetime_max_raw	Those attributes verbatim, as ISO 8601 with offset, e.g. 1970-01-22T20:05:00-05:00
date_source	Always frus:doc-dateTime in this snapshot. The code falls back to dateline when the attributes are missing
dateline_dates	Every <date> inside the document's datelines, verbatim: {text, when, from, to, notBefore, notAfter, calendar, type, ana}
places	<placeName> values from the datelines, e.g. ["Washington"]
source_note	Text of the first note[@type="source"], the archival provenance and classification. Missing for most pre-1945 documents
footnotes	All footnotes, in order: {id, n, type, text}. Footnotes are not part of text
correspondents	Elements in the heading that carry @type (from/to): {role, kind, text, ref}. ref points to glossary.entry_id when OH tagged it, which is about 12% of the time
person_ids, term_ids	Distinct persName/@corresp and gloss/@target ids used in the document. Join them to glossary on (volume_id, entry_id). The ids are local to a volume
page_start, pages	Printed page the document starts on, and page breaks inside it (<pb n>)
text	Full plain text: heading, dateline, body, attachments and signatures, without footnotes. Blocks such as paragraphs, list items and table rows go on separate lines; table cells are joined with |
n_chars	Length of text. Median 2,140; 191 documents exceed 100k characters; the largest has 1.26M

documents_tei: raw TEI for each document

volume_id, doc_id and tei_xml. tei_xml is the document's <div> serialised as standalone TEI XML, with namespaces declared. The rows are the same as in documents, in the same order and with the same shard file names, so you can join on (volume_id, doc_id) or zip the files row by row.

Why keep the TEI: it is lossless. It keeps tables, inline person and term tagging, <hi> emphasis, page breaks and footnote anchors, so you can re-render or re-parse without the 3.3 GB upstream repo. It is about half of the document bytes, so it lives in its own config and the default documents download stays around half the size.

volumes: one row per volume file

This table has all 744 files, including volumes that are not published yet. Columns:

• Titles: title (the complete title), series, subseries, subseries_n, volume_number and volume_title.
• People and publication: editors ({name, role}), publication_date, publisher, isbn13/isbn10/dospubno and source_description (e.g. "Born digital").
• Status: publication_status and status_changes. status_changes is the TEI revisionDesc/change list, {corresp, status, when, from, to, who, text}; in partially published volumes it gives each chapter's status.
• Coverage: content_date_*, administrations, production_priority and media_type.
• Counts: has_text, document_count (split into historical documents, editorial notes and other), pending_document_count, pending_chapter_ids and glossary_entry_count.
• Provenance: source_bytes, source_sha256, git_blob_sha and era_shard.

glossary: one row per entry in a volume's front-matter lists

list is persons (63,263 rows across 285 volumes) or terms (65,445 rows across 311 volumes, "Abbreviations and Terms"). The other columns are entry_id (e.g. p_RWP1, t_ACDA1), label (e.g. Acland, Anthony), description (e.g. Head of the Arabian Department, British Foreign Office, …) and the full item text.

sections: other front and back matter, one row per top-level section

Prefaces, "Sources", "About the Series", "Note on U.S. Covert Actions", errata, press releases, appendices, the printed back-of-book index, and for 19th-century volumes the President's annual message. Columns: section_id, part, type, subtype, title, url, contains_documents and text. text includes footnotes but leaves out any documents nested inside the section, which are documents rows. Two kinds of section are left out: the table of contents, which repeats path, and the persons and terms lists, which are in glossary.

Usage

from datasets import load_dataset
docs = load_dataset("taziksh/frus", "documents", split="train", columns=["volume_id", "doc_id", "date_min", "title", "text"])
vols = load_dataset("taziksh/frus", "volumes", split="train")
tei = load_dataset("taziksh/frus", "documents_tei", split="train")  # optional raw TEI

-- duckdb
SELECT volume_id, doc_id, date_min, title
FROM 'hf://datasets/taziksh/frus/data/documents/*.parquet'
WHERE date_min BETWEEN '1962-10-16' AND '1962-10-28' ORDER BY date_min;

How unpublished, partial and front-matter content is handled

• Unpublished volumes (planned, being-researched, being-cleared): upstream ships a header-only stub with a placeholder paragraph. They appear in volumes with document_count = 0 and has_text = false, so the table works as a status-of-the-series list.
• being-digitized (10): two general indexes (1861–99 and 1900–18) and eight 1958–63 microfiche supplements whose text is not encoded yet. They have metadata only.
• Partially published volumes (3: frus1969-76ve10, frus1977-80v27 and frus1981-88v16): only released documents become rows. In the last two, chapters still under declassification review are encoded upstream as div[@type="chapter-pending"] around div[@type="document-pending"] placeholders, which have a heading and no text. Those are counted in volumes.pending_document_count (696 in total) and pending_chapter_ids; they are not emitted as documents. frus1969-76ve10 has no placeholders. status_changes shows each chapter's status.
• Front matter: persons and terms lists go to glossary. Other sections go to sections. The table of contents is dropped. Documents that OH placed in front or back matter (106 of them, mostly in 1860s–70s volumes) are ordinary documents rows with part = front or part = back.

Known quirks and caveats

• Second editions are separate volumes. frus1951-54Iran and frus1951-54IranEd2 both have 375 documents, and frus1969-76ve15p2 (343 documents) sits alongside the revised frus1969-76ve15p2Ed2 (382). frus1977-80v09 is marked published but has an empty body, because its content lives in frus1977-80v09Ed2. Deduplicate on content if you need one copy per document.
• The pre-1950s volumes are different in kind. Volumes from 1861 to about 1945 ("Papers relating to …", 1861–1918) are mostly diplomatic correspondence:
  • most have no source notes (source_note is null for about 49k documents);
  • none has a persons or terms list (220 published volumes have no glossary);
  • headings are often of the form "Mr. Seward to Mr. Adams."
  • Some documents are very long reports or arbitration records, up to 1.26M characters.
• Non-English text. A few hundred pre-1946 documents are wholly or partly in French, Spanish or other languages, often printed with a translation alongside. This is a heuristic estimate; there is no language column.
• Very old dates are genuine. 15 documents dated before 1776 are historical exhibits, such as 17th-century charters, in the 1872 arbitration papers (frus1872p2v5) and a 1902 appendix (frus1902app2).
• Ids and markers inside the text. doc_id is unique within a volume, but (volume_id, doc_id) is the global key. Inline document identifiers, such as telegram numbers printed above the body, stay in text, e.g. a line reading 847.
• Sender and recipient tagging is sparse. correspondents comes from persName/gloss tagging in the heading, which OH applies unevenly. Most entries have no ref. The heading in title is the reliable form.
• Some ids may change. OH treats volume and document ids as canonical and stable. Person and term ids (p_…, t_…) are explicitly not stable, and are only meaningful within one volume.
• Plain text is a rendering choice. Footnote markers are dropped from text; footnotes[].n gives their numbers. Use tei_xml for anything layout-sensitive.
• Absent from OH's TOC. Five documents in front matter (frus1861 d1; frus1873p1v1 d1–d2; frus1874 d1–d2) are not linked from OH's generated table of contents, but are real document divs.

Reconciliation (this snapshot)

• Volumes: 744 upstream files in volumes/ → 744 volumes rows, with none missing and none extra.
• Documents: an independent regex count of <div … type="document"> over the raw XML gives 314,616, which equals the documents rows and the sum of volumes.document_count. There are no duplicate (volume_id, doc_id) keys, and the counts match for every volume.
• OH's generated tables of contents (frus-toc/*.xml) list every document id except the 5 front-matter documents above. They also list 697 ids that are not in this dataset: the 696 document-pending placeholders, and one div[@type="toc"] in frus1872p2v5 (d62a), which is not a document.
• Website spot check: documents frus1969-76v24/d176, frus1861/d51 and frus1981-88v16/d368 were checked against history.state.gov. All text lines, titles and footnotes match.

sync_state.json records the full reconciliation.

Licence and attribution

This dataset is released under CC0 1.0. The upstream LICENSE.md says:

As a work of the United States government, this project is in the public domain within the United States of America.

Additionally, we waive copyright and related rights in the work worldwide through the CC0 1.0 Universal public domain dedication.

and

Unless expressly stated otherwise, the person who associated a work with this deed makes no warranties about the work, and disclaims liability for all uses of the work, to the fullest extent permitted by applicable law. When using or citing the work, you should not imply endorsement by the author or the affirmer.

The repository README adds: "The files in this repository are in the public domain and may be copied and distributed without permission"; see https://www.state.gov/copyright-information/#copyright.

Upstream does not prescribe a citation format. The conventional form for a FRUS document, built from the volumes fields, is:

Foreign Relations of the United States, 1969–1976, Volume XXIV, Middle East Region and Arabian Peninsula, 1969–1972; Jordan, September 1970, ed. Linda W. Qaimmaqami and Adam M. Howard (Washington: Government Printing Office, 2008), Document 176. https://history.state.gov/historicaldocuments/frus1969-76v24/d176

For the source files, cite: U.S. Department of State, Office of the Historian, Foreign Relations of the United States TEI source files, https://github.com/HistoryAtState/frus (commit given above). This mirror is not affiliated with or endorsed by the Department of State.

Updates

Upstream publishes a semantic-release tag whenever OH posts new or corrected volumes. That happens roughly quarterly, with fixes in between. To stay in sync, build.py has two modes:

• uv run build.py full makes a shallow, bare clone of upstream (about 0.7 GB, no checkout). It streams each volume blob with git cat-file, parses the volumes with lxml (4 processes, about 2 minutes) and writes everything.

• uv run build.py update [--push] works in four steps:

  1. It reads the latest master commit and release from the GitHub API and compares the volumes/ tree's blob SHAs with sync_state.json.
  2. It downloads only changed volumes, from raw.githubusercontent.com at that commit.
  3. It rewrites only the affected era shards, carrying over the unchanged rows.
  4. It regenerates volumes, glossary and sections, sync_state.json and the stats block above.

  With --push it first pulls sync_state.json and the affected shards from the Hub, then uploads the changes in a single commit through huggingface_hub. If the parser version changes, it falls back to a full rebuild.

An incremental update produces byte-identical files to a full build; this was checked. Document shards are grouped into fixed eras by the volume id's first year (1861–1918, 1919–38, 1939–50, 1951–60, 1961–76, 1977–present), so an edited volume rewrites one documents shard and one documents_tei shard. sync-workflow.yml is an example GitHub Actions job (weekly check, needs an HF_TOKEN secret). A no-change run exits after two GitHub API calls.

Storage growth

The Hub keeps every version of every file. Its Xet storage deduplicates chunks of about 64 KiB, so a rewritten shard costs only the bytes that actually changed. Three things keep that small:

• Rows are ordered by volume, then by position in the volume, so one volume's rows are contiguous.
• Row groups never straddle two volumes; a large volume is split every 500 rows.
• Parquet content-defined chunking is on, with 64–256 KiB pages.

Measured with tools/dedup_estimate.py on three real upstream changes, replayed on this build. The figures are new bytes per sync, for the documents and documents_tei shards together:

change	no CDC, fixed 500-row groups	CDC defaults, fixed groups	CDC 64–256 KiB, volume-aligned groups (used)
edit inserting 47 documents into frus1952-54v09p1	96.7 MB	80.7 MB	3.0 MB
new volume frus1981-88v16 (stub to 88 documents)	9.1 MB	8.2 MB	1.5 MB
small fixes in 14 volumes across 2 eras (49e5401)	58.5 MB	52.4 MB	13.2 MB

With CDC at its default page sizes and row groups aligned to volumes, the costs were 6.5, 2.0 and 24.5 MB.

Over the past 12 months, 37 upstream commits touched volumes/. With a weekly sync that comes to 11 syncs, 60 changes to volumes that have documents, and 23 era-shard rewrites. At about 1–3 MB per changed volume, plus under 1 MB per sync for the small tables and state file, the projected growth is about 0.1 GB a year (range 0.07–0.2). The same changes without CDC and aligned row groups would cost about 0.8–1 GB a year. Volumes that are new rather than edited add their own size on top, typically a few MB each.