configs:
- config_name: arg_comp
data_files:
- split: train
path: arg_comp/train.json
- split: dev
path: arg_comp/dev.json
- split: test
path: arg_comp/test.json
- split: held_out
path: arg_comp/held_out.json
- config_name: arg_summ
data_files:
- split: train
path: arg_summ/train.json
- split: dev
path: arg_summ/dev.json
- split: test
path: arg_summ/test.json
- split: held_out
path: arg_summ/held_out.json
- config_name: case_retrieval
data_files:
- split: train
path: case_retrieval/train.json
- split: dev
path: case_retrieval/dev.json
- split: test
path: case_retrieval/test.json
- config_name: retrieval_corpus
data_files: retrieval_dataset.jsonl
- config_name: preprocessed
data_files: preprocessed.jsonl
task_categories: - summarization - text-generation language: - en tags: - legal-nlp - case-retrieval - legal-summarization - argument-completion size_categories: - 10K<n<100K
Dataset Card for BriefMe
Dataset Details
Dataset Description
Legal work involves reading, writing and reasoning about large volumes of text. For instance, in litigation, attorneys must search and interpret prior case law to construct convincing arguments to stake out a strong position. They present these to a judge in the form of a document called a brief.
BriefMe is a legal NLP corpus that presents three tasks on which to train and evaluate language models to assist in legal brief writing and argumentation:
(1) Argument Summarization: given a section of text from a brief, summarize the argument as a section heading for that part of the brief
(2) Argument Completion: given a set of arguments comprised of the section headings in a brief, complete the missing argument
(3) Case Retrieval: given a context window of text from the brief, retrieve the judicial opinion that supports the argument in the brief
Argument completion is a novel task that leverages the inherent structure of a brief, where briefs are split into multiple sections that constitute sub-arguments about the state of the law as interpreted by that attorney, and effective arguments fit together into a cohesive whole that logically walks the judge step-by-step to the attorney's preferred outcome.
BriefMe is built using legal briefs filed in cases argued in the Supreme Court of the United States.
- Language(s) (NLP): English
- License: CC-BY
Dataset Sources
- Repository: https://github.com/utahnlp/BriefMe
- Paper: BriefMe: A Legal NLP Benchmark for Assisting with Legal Briefs
Dataset Structure
Summarization data contains the following fields:
-file: str: The filename of the brief pdf that served as the original source of the text.
-text: str: The section of text from the brief to be summarized.
-reference: str: The header that corresponds to the text and also serves as a gold reference.
-judge_outcome: str: The output of the LLM judge's (GPT-o3-mini) rating of the reference.
-judge_score_verb: str: Text rating of the reference.
-judge_score: int: LLM judge score out of 5.
Completion data contains the following fields:
-file: str: The filename of the brief pdf that served as the original source of the text.
-text: str: The set of arguments represented by headings from the brief with one argument masked.
-reference: str: The missing argument heading from the brief.
-ToC_Avg_judge_score: float: LLM judge score out of 5.
Retrieval data contains the following fields:
-section_header: str: The header of the section that corresponds to the text that contains the context window.
-context: str: The text surrounding the masked citation that provides query context for the retrieval model.
-citation_value_orig: str: The case citation that points to the reference case in the standard legal form of <volume number> <reporter> <page number>.
-query_contains_other_citations: bool: Whether the retrieval query contains other citations within the context.
Retrieval corpus contains the following fields: - id: str: The citation as it would appear in legal text, composed of the volume, reporter abbreviation, and page (e.g., "538 U.S. 600") - cited_text: list or str: Either a list of strings containing the texts that match that citation in courtlistener, or if there is was no such match, a string ""
Preprocessed data contains the following fields: - filename: str: The unique filename of the brief document scraped from supremecourt.gov. - text: str: The raw text as extracted from the file using pypdf2. - docket_num: str: The docket number used by the U.S. Supreme Court to identify a case in its internal and public-facing systems. - court: str: The court that heard the case (all "SCOTUS") - arguments: str: The arguments as extracted from the table of contents. - cleaned_text_alt: str: The full text of the brief with non-ASCII and other encoding errors removed. - cleaned_content_alt: str: The longform text of the brief that has been split from the table of contents and other preceding material using regular expressions. - cleaned_args: list: A list of the extracted arguments from the table of contents. - sections_alt: list: A list of dictionaries with the argument as key matched to its section text in the brief as the value. - url: str: The url from where the pdf brief was downloaded. - arg_structure: json: A graph representing the relationships of the headers, with each header given its children and parents.
Data Splits
| Subset Name | Train | Dev | Test | Held-Out |
|---|---|---|---|---|
| arg_summ | 18,642 | 2,345 | 2,345 | 164 |
| arg_comp | 3,377 | 1,275 | 1,253 | 215 |
| case_retrieval* | 58,469 | 7,314 | 7,303 | - |
| Subset Name | Cases |
|---|---|
| retrieval_corpus | 24,525 |
*The size of the case_retrieval data splits is less than those reported in the paper because we applied additional filtering. Because of limitations with courtlistener, some examples cited to cases that were not contained in our retrieval corpus, and were removed from our experiments.
Dataset Creation
Curation Rationale
A model that could assist a legal professional in drafting briefs could offer notable productivity gains. Although there is a growing number of legal NLP datasets, none contain legal briefs in significant numbers or explicitly capture the logic of legal argumentation.
To this end, we present BriefMe, an annotated dataset of Supreme Court of the United States (SCOTUS) briefs with annotations to support three tasks. We construct BriefMe from briefs hosted on the Supreme Court's website.
We formulate the tasks based on the structure of SCOTUS briefs. This structure follows a convention where each section and subsection state a concise version of the arguments that are contained in that section, with each section building toward the ultimate argument. Leveraging this structure, we construct a dataset where each section or subsection is accompanied by a section heading, which can serve as a built-in gold reference for a summarization task. Taken together, the headings capture the step-by-step argument of the brief, and form the basis for a novel argument completion task.
Source Data
Briefs were scraped from www.supremecourt.gov. The judicial opinion retrieval corpus is compiled using the Court Listener API.
Data Collection and Processing
Review the related paper for details on data collection and processing.
Who are the source data producers?
Briefs were produced by attorneys arguing cases in front of the Supreme Court. Judicial opinions are produced by U.S. judges.
Annotations
Annotation process
Headings for the argument completion and summarization tasks include a rating from an LLM judge that was calibrated against human annotators with legal training retained through Qualtrics. These ratings can be used to filter the data for a higher quality set. See the related paper for additional details.
Personal and Sensitive Information
The data may contain personal information, although all data is from publicly filed court documents or judicial opinions, so it may not be considered "private."
Bias, Risks, and Limitations
[More Information Needed]
Citation
BibTeX:
@inproceedings{woo2025BriefMe, author = {Woo, Jesse and Hashemi Chaleshtori, Fateme and Marasovi{'c}, Ana and Marino, Kenneth}, title = {BriefMe: A Legal NLP Benchmark for Assisting with Legal Briefs}, booktitle = {{ACL} '25: Findings of the Association for Computational Linguistics, Vienna, July 27 - August 1, 2025}, year = {2025}, }
APA:
Jesse Woo, Fateme Hashemi Chaleshtori, Ana Marasović, and Kenneth Marino. 2025. BriefMe: A Legal NLP Benchmark for Assisting with Legal Briefs. In Findings of the Association for Computational Linguistics: ACL 2025, Vienna, Austria. Association for Computational Linguistics.
Dataset Card Contact
Jesse Woo: jw4202@columbia.edu