alea-institute/alea-docket-chain

ALEA Docket Chain

ALEA Docket Chain is a set of 43,463 short reading tasks over real U.S. federal court dockets (42.7 million tokens in the alea-v2b-sw32k tokenizer), released in two parts (see Releases). Each document shows a contiguous excerpt of a docket and asks questions that need two or three steps of reference resolution: from an order to the motion it decides, and from that motion to who filed it, what it asked for, when it was filed, or how it was decided.

The data was built by the ALEA Institute with deterministic code only (no language model wrote any text or label).

Releases

release	file	documents	tokens	what is new
v1	data/train-00000-of-00001.parquet	26,922	26.3M	first release
v2	data/train-v2-00000-of-00001.parquet	16,541	16.4M	new documents from dockets that v1 did not use
total	data/train-*.parquet (the train split loads both files)	43,463	42.7M

v2 is built with the same code, rules, filters and review process as v1 from the same docket collection. It uses only dockets that v1 did not use, either as a target, as a worked example or as a candidate that was later screened out. Its rows are new; the v1 file is unchanged (git tags v1 and v2 mark the two releases). v1 used up most of the dockets that support the balanced side and ruling questions, so v2 has more title and date questions and fewer 3-step questions (see Composition).

Source

• Docket text comes from the KL3M copy of the public CourtListener / RECAP docket records of U.S. district courts and courts of appeals. Docket entries are shown verbatim and in their original order (about a third of the excerpts list them newest first). Nothing in an entry is edited.
• The header gives the court, case name and docket number, and sometimes the assigned judge, the filing date and a party list. The party list holds only names and roles (Plaintiff, Defendant, Petitioner, Respondent) from the docket's own party records, and it is shown only when every party of the case has one clear role. No attorney contact data is used.
• Bankruptcy courts and magistrate (-mj-) dockets are not used. Excerpts with e-mail addresses, phone numbers, Social Security numbers, private street addresses, P.O. boxes or inmate register numbers are dropped. Case names and party names are public record and are kept.
• v1 covers 93 courts and v2 covers 94. In each release no court is more than 5% of the documents.

Task

Every document is an instruction, the docket excerpt and one to three questions, followed by the answer. Question types:

type	example question	answer	steps
side	Which side filed the motion that entry 36 granted?	The plaintiffs.	2 (role stated in the motion entry) or 3 (motion entry gives a name; the party list gives its role)
name	Who filed the motion that entry 22 denied?	the party name, copied from the entry	2
what	Which motion was granted in entry 45? Give its docket title.	the motion title as docketed	2
date	When was the motion that entry 24 ruled on filed?	the filing date shown for the motion's entry	2
outcome	Was the motion the defendant filed on 08/04/2010 granted or denied?	Denied (entry 80).	2 or 3
order	Which entry ruled on the motion the plaintiff filed on May 14, 2018?	Entry 29.	2 or 3
target	The motion that entry 75 decided refers to an earlier motion. Which side filed the earlier motion?	The petitioner.	3
opp	Which side opposed the motion that entry 130 denied?	The defendants.	3

Answers are short (a side, a name, a title, a date, an entry number or a ruling), sometimes as one sentence that cites the entry numbers. Presentations: plain text (40%), question/answer markers (30%) and chat turns (30%) in both releases; 6% (v1) and 5.5% (v2) of documents carry one or two worked examples from other dockets whose answers never match the target's.

How the answers were derived

• An order is used only if it decides exactly one numbered motion in a single "granting / denying ... N Motion" clause, and no other order rules on that motion. Orders that rule on several motions are dropped.
• The filer of a motion is taken from the docket's own party records: the whole "by ..." clause must consist of party names of one side. Joint, stipulated, consent, agreed and "on behalf of" filings are not used, nor are titles whose possessive role ("Government's Motion") disagrees with the bound filer.
• A fact is used only when nothing else in the whole docket can change it: chains are dropped when a later entry refers to the motion or the order with vacate, reconsider, withdraw, strike, moot, amend, modify, continue, appeal or "entered in error" language, when the motion was refiled as an amended or renewed motion, or when either entry is marked in error, stricken, deficient or terminated.
• Every answer string (name, title, date, role word or party-list entry) occurs literally in the shown excerpt.

Balancing against shortcuts

Figures are v1 / v2.

• The nearest motion before an order is usually the one it decides. Questions were sampled so that "answer from the nearest motion above the order" is right only about 40% of the time for names, titles and dates (v2: 40.3 / 40.0 / 39.8%), and about 52% / 53% for sides.
• Plaintiff and defendant answers are balanced 50.5 / 47.7 (v2: 49.8 / 50.2; petitioner / respondent 1-2%); "granted" is 55% / 55% of ruling answers.
• A bag-of-words model that sees only the question does no better than the majority class (side: -0.8 / -2.6 points; granted vs not granted: -1.6 / -2.7 points).
• When an order's own text names the movant's role ("Defendants' Motion"), the side question becomes one step shorter; such documents are capped at 15% of side questions.
• 5.6% (v1) and 7.5% (v2) of documents are twins: two documents over the same excerpt with the same question wording and different answers.

Composition

	v1	v2
documents	26,922	16,541
questions	31,613 (2-step 65%, 3-step 35%)	18,780 (2-step 91%, 3-step 9%)
question types	date 25%, outcome 23%, what 20%, side 16%, name 10%, order 5%, opp and target 1%	date 50%, what 30%, name 8.5%, outcome 8%, side 2.6%, order 1%, opp and target 0.3%
courts (largest share)	93 (5.0%)	94 (5.0%)
documents with a party list	61%	43%
tokens per document	mean 977, 99th percentile 2,514, max 3,559	mean 993, 99th percentile 2,559, max 3,749

Review

For v1, three reads checked every answer against its excerpt: two in-house reads of 30 documents each and an independent review of 60 randomly drawn documents that saw only the rendered text (the in-house reads found three errors: an order ruling on three motions, a motion title cut at "as to", and a pro se party's mailing address in an order). The independent review found 2 errors in 60 (3.3%): an order ruling on several motions, and a motion title cut at a parenthetical. These error classes (including orders that decide cross-motions in one line), and the minor issues found (stray filer words in titles, contradictory title roles, phone-like numbers, and side hints in order text), were removed from the data by dropping documents.

v2 review

v2 had its own review on documents drawn only from v2. A builder read of 20 documents found no errors. An independent review of 60 documents (rendered text only) found 3 errors in round 1 (5%): two motion titles that ran on into fee or receipt text or were cut inside a parenthesis, and one document where one question's wording gave away another question's answer. All three error classes were removed by dropping documents, along with two related issues: a motion entry holding two motion events, and a single document asking both the name and the side of the same filer. A second independent review of 60 fresh documents (more title and multi-question documents) found 2 errors (3.3%): a "which entry ruled on the motion filed on date D" question where the docket shows a second motion on that date, and an "opposition" that was a notice of non-opposition. Both classes were then also removed by dropping documents. No text was edited in either release.

Columns

• id: document id
• text: the full rendered document (instruction, excerpt, questions and answer)
• family: always docket_chain
• kind: type of the first question
• source: CourtListener court id (e.g. txsd, cand)
• gold: JSON list with one record per question: question text, gold answer, order entry (O), motion entry (N), referenced entries where relevant (K, R), number of steps, answer form
• presentation: plain, qa or chat
• group_id: documents sharing an excerpt (twins) share a group
• hops: largest number of steps among the document's questions
• n_tokens: length in alea-v2b-sw32k tokens including the two document boundary tokens

License

CC BY 4.0. The docket text is U.S. federal court public record.