codexpublicus/stare-citations

stare.pub citation spans

US court decisions from stare.pub, each with its citation spans labelled by every source we have and tiered by how far each label can be trusted. It is built for training and evaluating citation extraction and resolution. stare.pub (https://stare.pub) is the source and stays authoritative: this release is v2026.10.24 at https://data.stare.pub/datasets/stare-citations/v2026.10.24/, and this repository mirrors it, one tagged revision per release.

from datasets import load_dataset

tiered = load_dataset("codexpublicus/stare-citations", "tiered", revision="v2026.10.24")
review = load_dataset("codexpublicus/stare-citations", "review", revision="v2026.10.24")

Configs

• tiered (default): one row per decision, 10,000 rows. Each row has the decision's text (text.content), where stare.pub serves it, its strata and its citations: every citation in at least one tier, 223,412 in all.
• review: one row per citation in no tier, 45,622 rows, with 160 characters of context either side, every source's label, what the sources disagree on (kinds) and its place in the review order (rank, pattern, exemplar). These are the citations a person rules on; a ruling puts a citation in tier 0 in the next release.

The split is named train only because the Hub needs one. It holds the whole release. This dataset has no training, validation or test split, and no hold-out flag: splits, evaluation sets and exclusions are the consumer's choice. citation-bench's vectors come from 113 of these decisions; a consumer that evaluates against citation-bench can find them by each vector's source.

Sources and labels

Every label names its source, and every citation lists which sources found it (found_by), which could have and didn't (missed_by), what they disagree on (disagreements) and whether a person ruled on it (human_ruled).

• citare@0.7.3: citare's output, built from the v0.7.3 tag.
• eyecite@2.7.8: Free Law Project's eyecite, run with PYTHONHASHSEED=0 because its output varies with Python's hash seed.
• cap-cites_to: the Caselaw Access Project's own cites_to list for a CAP case (Harvard LIL's extraction), placed in the text by string search. It has no offsets of its own, and since LIL co-developed eyecite it is treated as possibly eyecite-related.
• curated: a person's label, from citation-bench's vectors or a ruling on the review queue.

Each engine resolves a case citation to a decision through stare.pub's citation index (target, keyed by stare.pub's key: cap/<slug>/<folder>/<file> or cl/<cluster>/<hash>). Spans are [start, end) in Unicode code points over text.content, the decision's opinions joined with a blank line.

Ground truth is curated labels verified against an official source or ruled on by the owner, and nothing else. eyecite's output is never ground truth, and neither is any engine's.

Tiers

A citation's tiers lists every tier it is in and tier the lowest. The tier belongs to the citation (the span and every source's reading of it), not to one source's label. The rules are the owner's rulings D87, D91 and D95 (2026-10-23) and D107 (2026-10-24).

Tier	Name	A citation is in it when	Citations
0	verified	a curated label on it is ground truth: an owner-verified vector, a citation-bench vector whose statute claims an official source confirmed, or an owner ruling on the queue	33
1	consensus	every source that could have found it found it at the same characters, with the same type and parsed fields, and every source that names a case names the same one, the engines always among them	215,904
2	metadata-confirmed	the case the sources resolve it to lists the cited volume, reporter and page among its own citations, in CAP's case metadata (cap-metadata) or CourtListener's cluster citations (cl-metadata)	205,381
	in a tier (tiered)		223,412
	in no tier (review)		45,622

Overlaps: 1 and 2 197,901; 1 alone 17,998; 2 alone 7,480; 0 alone 28; 0 and 1 5.

• Consensus counts cap-cites_to only for a full case citation in a CAP decision that has a cites_to list; CourtListener records have none, so their consensus is the two engines. cites_to usually lists one citation of a parallel run, so a citation it misses because it listed the other half (missed_by:cap-cites_to:parallel) needs only the engines. A short form, supra or Id. resolves to its antecedent's case.
• Metadata confirmation confirms the resolution, not where the span starts and ends. metadata_confirmed.sources names every list that has the citation, so a consumer can drop either: cap-metadata confirms 175,263 citations and cl-metadata 201,049. CourtListener's cluster citations come from its ingest paths (court-website scrapers, the CAP import, the Supreme Court Database, a Westlaw dataset), each of which parses a citation string with eyecite but none of which extracts citations from opinion text. A short form is confirmed when its antecedent is.
• A CourtListener record's own citation (D107 (2)). Many CourtListener records open with their own citation (80 P.3d 171 (2003), [Cite as State v. Armengau, 2018-Ohio-4299.]), and stare.pub's index can hold another cluster at the same volume and page. From the top of a record's text, every citation that is one of the record's own citations, up to the first that isn't, resolves to the record itself, as do short forms that refer back to it. It is in tier 1 (and tier 2 when the record's citations list it) when the engines agree it is the record, and otherwise a review item.
• Bare section signs (D107 (1)). A section sign with no section number (§, §§, (§), which both engines label unknown, is not a citation: it is in neither config. 26,835 were dropped from this release, and the release manifest.json counts them under counts.dropped.

Tags

tags say how a citation stands where its tier alone doesn't. A tag is not a tier.

• span-only (32,924 citations): cap-cites_to matched the span of a case citation but linked no case, or several (a page of memorandum decisions such as 321 U.S. 790), so it abstained on the case. Such a citation is in tier 1 when the engines agree on the case (D95). A consumer that wants every source's case can drop these from tier 1.
• no-case-to-resolve (14,581 citations): a statute, journal or other citation of no case, in tier 1 on span, type and fields alone (D91 (3)). A case citation that resolves to nothing is not one of these: it is in neither tier 1 nor tier 2 and is a review item.
• own-citation (1,532 citations): a CourtListener record's own citation at the head of its text, resolved to the record itself (D107 (2)). 1,524 are in tiers 1 and 2; 8 the engines don't agree on are review items (6 Tex. Bus. neutral citations only citare finds, and 2 S.Ct. citations where the engines' spans differ by the year). For 574 of them the index names a different decision at the same volume and page.

Sampling

10,000 decisions: 8,000 Caselaw Access Project cases and 2,000 CourtListener records (140,719,086 characters), drawn with seed 86 from a frame of stare.pub's 6,920,598 CAP cases in 40,622 volumes and the 900,342 records of CourtListener's 2026-09-30 bulk export, against stare.pub index v2026.10.23. CourtListener records are about 11.5% of stare.pub's decisions and 20% of the sample, so their tiers, which rest on two engines, can be read on their own.

The draw is stratified by source, decade, jurisdiction (each state and territory, each federal circuit) and court level (highest, intermediate, trial, other): every decision a citation-bench vector comes from (113); then floors of 75 decisions a decade, jurisdiction and level, using at most 60% of each source's quota (5,887); then a uniform fill (2,000); then rare kinds of citation topped up from a 20% reserve (2,000): Id., supra, short forms, statutes per jurisdiction, nominative reporters, pre-1850 decisions, neutral and slip citations, and OCR damage. A decision counts toward a statute bucket only when citare or eyecite labels a statute in it. The statute top-ups took the whole reserve, so no uniform fill followed them. The floors and top-ups oversample small strata and rare kinds, so the sample as a whole is not representative of the corpus; each row's stratum.reason says how it was drawn (curated, floor:<dimension>=<bucket>, fill, kind:<bucket>). Decades before the 1790s are short, because the corpus is.

Known gaps

• Statutes. The engines miss common state statute forms, so the statute buckets the draw aimed at 75 decisions each are thin where those forms are used: the thinnest are Louisiana (6 decisions with a statute label), Kentucky (8), Florida (10), Oregon (12), Missouri and Ohio (13), and Illinois and Michigan (17); 27 of the 55 state and territory buckets reach 75. Neither engine labels a code's own abbreviation without § (citare#115: NRS 284.383, ORS 163.375, R.C. 2929.20), and they miss the section before the code (citare#116: Section 54.28, Fla.Stat., F.S.A.), the section spelled out or with a comma or year (citare#117: Labor Code section 3602), and session laws, dated revisions and civil code articles (citare#118: Ill. Rev. Stat. 1987, ch. 121, par. 4-501, La. C.C. art. 1931). A citation no source finds is in neither config.
• Unresolved case citations. A case citation to a reporter volume stare.pub doesn't index, to a page with no case, or in a form the index doesn't key (neutral and WL citations) resolves to nothing, so it is in neither tier 1 nor tier 2 and is a review item: 32,470 of the 45,622.
• Tier 0 is small: 33 citations. Ground truth grows as the review queue is ruled on.
• Agreement is not accuracy. Tier 1 is where independent sources agree; engines that share a blind spot agree on it too, and eyecite and cap-cites_to may not be independent.
• CourtListener text comes through stare's conversion of html_with_citations, which strips CourtListener's own citation anchors, so those records carry no span labels from CourtListener.

Format

schema_version 4.1.0 (schema stare-train/4), SemVer: a field added is a minor version; a field renamed or removed, or a meaning changed, a major one. Pin it. Each citation has a stable id, its decision's key and a hash of its span and characters, so the same citation in the same text has the same id in every release and releases can be diffed by it; a review item's citation is that id. The format is documented in stare-builder's pipelines/train-export/README.md.

The release directory has the same rows as JSON lines (jsonl/tiered/records.jsonl, jsonl/review/queue.jsonl, the export's files byte for byte), the export's own manifest.json, stats.json and stats.md under export/, and a release manifest.json that pins the stare.pub index build and reporters.json, the CourtListener export date, the engine versions and commits, the citation-bench commit and statute verdicts, the tier rules, the sampling seed and frame, counts by tier, the dropped bare section signs and the CourtListener records' own citations, and every file's size and SHA-256. Input files are named by file name and SHA-256, never by a path on the machine that built the release. CHANGELOG.md says what changed from release to release.

Licence and attribution

The Hugging Face licence field is cc-by-nc-nd-4.0. The parts of this release have these terms:

• Labels and tiers: CC BY-NC-ND 4.0, with stare.pub (https://stare.pub) as the source. stare.pub states that everything it produces, other than the Caselaw Access Project's case files, is CC BY-NC-ND 4.0, crediting stare.pub as the authoritative source.
• Caselaw Access Project text: the CAP case text and metadata are CC0 1.0, the terms the Caselaw Access Project published them under.
• CourtListener text: the CourtListener records carry the Public Domain Mark 1.0 that CourtListener puts on its bulk data, with credit to CourtListener and its publisher, Free Law Project, from the 2026-09-30 bulk export. Each record states this in case.cl.license and case.cl.attribution.

stare.pub's index build names its sources the same way: the Caselaw Access Project data (CC0 1.0) and CourtListener's bulk data (Public Domain Mark 1.0, Free Law Project).