docketx/us-dockets

US Dockets — the case-level record

Code & tools: github.com/docketx — legal-scrambler pseudonymises a case file on your own hardware before a frontier model sees it; claude-for-legal is the Claude Code plugin (docketx-open-law) that loads these datasets and checks citations against them.

34,290,148 docket records from 2,289 U.S. courts: the case, the court, the number, the dates, the judge. This is the index of American litigation — what was filed, where, when and by whom it was decided.

Source & credit — Free Law Project / CourtListener

Every docket in this dataset comes from the Free Law Project / CourtListener bulk export of 2026-06-30. CourtListener gathered, cleaned, deduplicated and published that record; Free Law Project — the 501(c)(3) non-profit behind it, building free public access to American law since 2010 — pays for it. We reshaped and sharded; they did the hard part. Go to the source: <https://www.courtlistener.com>.

Free Law Project is a non-profit, and it is worth supporting: <https://free.law/donate/>. If this data is useful to you, credit Free Law Project and CourtListener too, and consider going straight to the source: CourtListener runs its own API and MCP server.

What a docket row IS — and what it is NOT

A docket row is metadata about a case. Case name, court, docket number, filing date, termination date, cause of action, nature of suit, assigned judge, and the CourtListener URL for the case.

A docket row is NOT the filings. There are no complaints, no motions, no briefs, no orders, no exhibits, no opinions here. In the federal courts those documents live behind PACER, which charges per page; the RECAP Archive at CourtListener holds the subset that users have paid for and donated. Nobody has them all — not Free Law Project, not us, not any commercial vendor. Anyone who tells you their corpus contains "all court filings" is describing something that does not exist.

A docket is not a "case" in the sense a citation is. One dispute can carry several dockets (trial, appeal, consolidated, MDL member cases), and a docket can be opened and closed without producing a single published decision. Do not quote a docket count as a count of cases, of lawsuits, or of decisions. For decisions, use `docketx/us-caselaw`.

A docket row is not the whole docket sheet. The chronological list of entries — "12/03 Motion to dismiss filed" — is a separate CourtListener table and is not in this dataset.

Privacy — what we removed, and why

Docket metadata names private individuals. CourtListener carries a blocked flag on each docket, which it sets when the record should be kept out of search-engine indexes; in this export the flag is applied at scale to consumer bankruptcy, where the party is an ordinary person and the subject is their insolvency.

We excluded every `blocked` docket: 37,369,473 rows. Publishing them here — in a dataset that is indexed, bulk-downloadable and read by model trainers — would defeat exactly the protection the source applied. They are not hidden: they remain in Free Law Project's own bulk export and on CourtListener, for anyone with a reason to have them.

We checked every column of the source table for fields that should not be redistributed. There are no Social Security numbers, dates of birth, addresses or juvenile-matter markers in the schema; the only privacy control the source carries is the blocked flag described above, and we honoured it. We also dropped CourtListener's internal operational columns (view_count, Internet Archive upload paths and retry counters, index timestamps), which say nothing about the case.

Malformed rows at source

18,026 rows of the source CSV carry an unescaped " inside a text field, which shifts every following column — a party's mailing address lands in the docket-number column, and so on. We detected these by an impossible court_id — values like f, 1988, 0, Inc. — and dropped them rather than publish shifted values. They are not in any config and not in this repo: because the shift also scrambles the blocked column, we cannot tell which of them the source had marked private, so they are excluded on that ground too. The check is reproducible from the source export: a row whose court_id is absent from courts-2026-06-30.csv.bz2 is one of these.

Composition — measured, per tier

Every number on this page was counted from the files in data/ — none is estimated, and none is taken from any upstream total. Read from the source export: 71,677,647 rows.

The record is federal-heavy, and that is a property of the world, not of this dataset: PACER gives CourtListener a machine-readable feed for every federal court, while most state trial courts publish no bulk docket data at all. A state absent from a count is one with no open feed, not one withheld.

The viewer, and why two configs preview partially

Hugging Face converts each config to parquet up to a 5 GB in-memory ceiling, measured here rather than assumed: federal-district stopped at 5,001,023,741 bytes and all at 5,003,462,887. So in the browser those two show a partial preview — federal-district previews 7,490,694 of its 23,795,525 rows, all previews 7,596,221 of 34,290,148.

Nothing is missing from the data. load_dataset and a plain download read every shard; the cap applies only to the browsable parquet preview. The other eight configs convert completely, and every one of their converted row counts matches the count measured on our own files exactly: state 6,841,279 · federal-special 1,605,763 · federal-appellate 1,341,708 · scotus 589,547 · territorial 63,006 · federal-bankruptcy 27,619 · other 23,563 · tribal 2,138.

We kept the tier sharding rather than splitting the federal district courts further, because the alternatives — by circuit, by state — are not in the source table (CourtListener gives every district court the same parent_court_id, usdistct), and we do not add facts the source does not carry. Files are held at about 100 MB each, 40 in total, which is the size the converter handles comfortably.

Format

Rows follow docketx record v1, one docket per line, every column a string:

id, doc_type (docket), jurisdiction, court, court_id, tier, title, docket_number, date_filed, date_terminated, cause, nature_of_suit, assigned_to, source, license, retrieved_at, extra.

extra is a JSON string — parse it with json.loads — carrying everything else the source exposed: the case-name variants, docket_number_core, pacer_case_id, jury demand, MDL status, appellate fee status, the cert/argument dates, the panel and referral strings, and the CourtListener judge ids (assigned_to_id, referred_to_id) which join to `docketx/us-judges`.

jurisdiction is the two-letter state code for state courts and us for federal ones; use court_id and court to narrow further (e.g. nysd = District Court, S.D. New York).

Load it

from datasets import load_dataset

ds  = load_dataset("docketx/us-dockets")                       # every tier
fd  = load_dataset("docketx/us-dockets", "federal-district")   # one tier
tx  = ds["train"].filter(lambda r: r["jurisdiction"] == "tx")  # one state

Provenance and license

U.S. court records are works of government and carry no copyright (Banks v. Manchester, 128 U.S. 244 (1888); Georgia v. Public.Resource.Org, 590 U.S. 255 (2020)). This packaging is released under CC0 1.0. Source: CourtListener bulk export of 2026-06-30, Free Law Project — https://free.law. This dataset redistributes public-record metadata; it adds no annotation and asserts no rights.

Part of the DocketX open legal corpus, alongside docketx/us-caselaw (10.8M opinions), docketx/us-judges, docketx/court-rules, docketx/us-statutes and docketx/oral-arguments-us.