Datasetpublic

StepUpLaw/disputed-terms-national

Disputed Contract and Instrument Terms in US State Appellate Courts One row per (term, decision) pair: a word or phrase a court quoted as the disputed language in a decision containing an ambiguity holding, across the state appellate courts of all 50 states and DC. Terms are extracted from the whole opinion by cue-anchored quotation ('the term "X"', '"X" as used in'), classified into functional drafting categories, and linked to the holding by three graded levels of evidence.… See the full description on the dataset page: https://huggingface.co/datasets/StepUpLaw/disputed-terms-national.

sourceHugging Facecc-by-4.0updated 21d agoView on Hugging Face

0likes84downloads

Dataset Card

Disputed Contract and Instrument Terms in US State Appellate Courts

One row per (term, decision) pair: a word or phrase a court quoted as the disputed language in a decision containing an ambiguity holding, across the state appellate courts of all 50 states and DC. Terms are extracted from the whole opinion by cue-anchored quotation ('the term "X"', '"X" as used in'), classified into functional drafting categories, and linked to the holding by three graded levels of evidence.

286,846 rows. Licence CC BY 4.0. Not legal advice.

Where this comes from


Canonical record	https://doi.org/10.5281/zenodo.22247377
Code and methodology	https://github.com/stepuplaw/legal-empirics
Research page	https://stepuplaw.com/research/
Author	Kevin D. Klagge, ORCID 0009-0002-1385-8498
Source corpus	CourtListener bulk export, snapshot 2026-06-30

The DOI above identifies the code, which is a different object from this dataset. Cite the code when you are describing the method and cite this dataset when you are using the numbers.

Columns

Column	Type	Meaning
`term`	string	the disputed language, lowercased and normalised, as the court quoted it
`category`	string	functional drafting class: nexus, degree, temporal, scope, modal, role, succession, property, event, conduct, condition, mental, uncategorised
`source`	string	the instrument the words sit in: testamentary, deed, insurance, contract, statute, constitution, uncertain
`posture`	string	found	rejected	alleged	uncertain -- whether the court held the language ambiguous
`link`	string	direct	proximate	inferred -- strength of the link between this term and the holding; inferred rows must not be pooled with the other two
`oid`	integer	CourtListener opinion id
`cid`	integer	CourtListener cluster id
`court`	string	CourtListener court id
`state`	string	two-letter state postal code
`year`	integer	year the decision was filed
`statement`	string	the row written as one self-contained English sentence, so it can be retrieved, quoted and checked on its own

Every row carries a statement column, which is the row written as one self-contained English sentence. A row of codes can be downloaded but not retrieved or quoted, and the sentence is what makes each row usable on its own.

How it was built

Retrieval and extraction are deterministic code over a local corpus of 10.8M US judicial opinions. Classification uses rules written against a hand-coded sample that ship with their measured accuracy, so the error rate is reported rather than assumed. Every study states its exclusion funnel with counts, because silent filtering is the commonest defect in research on opinions and it is invisible in the result.

datapackage.json carries the Frictionless schema, croissant.json the MLCommons Croissant description, and dataset.jsonld the schema.org form.

Limits

This is exploratory. The coded samples behind it were coded once, so it supports a described pattern rather than a measurement. Inter-annotator reliability has not been established.

Published appellate opinions are not disputes. Most disputes settle, most settlements are unpublished, and appellate coverage varies by court and decade. Any rate here is a rate among decisions that reached an appellate court and were published, which is not the same population a drafter cares about.

Read the study's own limitations section in the repository before quoting a number. Each one names the specific threats to its own validity, including the ones that are unflattering.