Hugging Face's logo

Datasets:
harvard-lil
/
cold-cases

Modalities:
Tabular
Text
Formats:
parquet
Languages:
English
Size:
100K - 1M
Tags:
united states
law
legal
court
opinions
Libraries:
Datasets
Dask
License:
Dataset card Viewer Files Community
12
cold-cases / README.md
MatteoCargnelutti's picture
MatteoCargnelutti
"Table of contents" appears to be a reserved word in README.md
85034fa
preview code
|
raw
history blame
7.95 kB
metadata
license: cc0-1.0
language:
  - en
tags:
  - united states
  - law
  - legal
  - court
  - opinions
size_categories:
  - 1M<n<10M

Collaborative Open Legal Data (COLD) - Cases

COLD Cases is a dataset of 8.3 million United States legal decisions with text and metadata, formatted as one JSON object per decision. The total dataset size is approximately 104GB of uncompressed JSON.

This dataset exists to support the open legal movement exemplified by projects like Pile of Law and LegalBench. A key input to legal understanding projects is caselaw -- the published, precedential decisions of judges deciding legal disputes and explaining their reasoning. United States caselaw is collected and published as open data by CourtListener, which maintains scrapers to aggregate data from a wide range of public sources.

COLD Cases reformats CourtListener's bulk data so that all of the semantic information about each legal decision (the authors and text of majority and dissenting opinions; head matter; and substantive metadata) is encoded in a single JSON object per decision, with extraneous data removed. By consolidating the data engineering for preprocessing caselaw in an open source pipeline maintained by the Harvard Law School Library, we ensure that downstream machine learning and natural language processing projects can use consistent, high quality representations of cases for legal understanding tasks.

Prepared by the Harvard Library Innovation Lab in collaboration with the Free Law Project.

Links

Summary

Formats

We've released this data in two different formats:

JSON-L or JSON Lines

This format consists of a JSON document for every row in the dataset, one per line. This makes it easy to sample a selection of the data or split it out into multiple files for parallel processing using ordinary command line tools such as head, split and jq.

Just about any language you can think of has a ready way to parse JSON data, which makes this version of the dataset more compatible.

See: https://jsonlines.org/

Apache Parquet

Parquet is binary format that makes filtering and retrieving the data quicker because it lays out the data in columns, which means columns that are unnecessary to satisfy a given query or workflow don't need to be read.

Parquet has more limited support outside the Python and JVM ecosystems, however.

See: https://parquet.apache.org/

☝️ Go back to Summary

File structure

Both of these datasets were exported by the same system based on Apache Spark, so within each subdirectory, you'll find a similar list of files:

  • _SUCCESS: This indicates that the job that built the dataset ran successfully and therefore this is a complete dataset.
  • .json.gz or .gz.parquet: Each of these is a slice of the full dataset, encoded in JSON-L or Parquet, and compressed with GZip.
  • Hidden .crc files: These can be used to verify that the data transferred correctly and otherwise ignored.

☝️ Go back to Summary

Data dictionary

Partial glossary of the fields in the data.

Field name Description
judges Names of judges presiding over the case, extracted from the text.
date_filed Date the case was filed. Formatted in ISO Date format.
date_filed_is_approximate Boolean representing whether the date_filed value is precise to the day.
slug Short, human-readable unique string nickname for the case.
case_name_short Short name for the case.
case_name Fuller name for the case.
case_name_full Full, formal name for the case.
attorneys Names of attorneys arguing the case, extracted from the text.
nature_of_suit Free text representinng type of suit, such as Civil, Tort, etc.
syllabus Summary of the questions addressed in the decision, if provided by the reporter of decisions.
headnotes Textual headnotes of the case
summary Textual summary of the case
disposition How the court disposed of the case in their final ruling.
history Textual information about what happened to this case in later decisions.
other_dates Other dates related to the case in free text.
cross_reference Citations to related cases.
citation_count Number of cases that cite this one.
precedential_status Constrainted to the values "Published", "Unknown", "Errata", "Unpublished", "Relating-to", "Separate", "In-chambers"
citations Cases that cite this case.
court_short_name Short name of court presiding over case.
court_full_name Full name of court presiding over case.
opinions An array of subrecords.
opinions.author_str Name of the author of an individual opinion.
opinions.per_curiam Boolean representing whether the opinion was delivered by an entire court or a single judge.
opinions.type One of "010combined", "015unamimous", "020lead", "025plurality", "030concurrence", "035concurrenceinpart", "040dissent", "050addendum", "060remittitur", "070rehearing", "080onthemerits", "090onmotiontostrike".
opinions.opinion_text Actual full text of the opinion.
opinions.ocr Whether the opinion was captured via optical character recognition or born-digital text.

☝️ Go back to Summary

Notes on appropriate use

When using this data, please keep in mind:

  • All documents in this dataset are public information, published by courts within the United States to inform the public about the law. You have a right to access them.
  • Nevertheless, public court decisions frequently contain statements about individuals that are not true. Court decisions often contain claims that are disputed, or false claims taken as true based on a legal technicality, or claims taken as true but later found to be false. Legal decisions are designed to inform you about the law -- they are not designed to inform you about individuals, and should not be used in place of credit databases, criminal records databases, news articles, or other sources intended to provide factual personal information. Applications should carefully consider whether use of this data will inform about the law, or mislead about individuals.
  • Court decisions are not up-to-date statements of law. Each decision provides a given judge's best understanding of the law as applied to the stated facts at the time of the decision. Use of this data to generate statements about the law requires integration of a large amount of context -- the skill typically provided by lawyers -- rather than simple data retrieval.

To mitigate privacy risks, we have filtered out cases blocked or deindexed by CourtListener. Researchers who require access to the full dataset without that filter may rerun our pipeline on CourtListener's raw data.

☝️ Go back to Summary