system's picture
system HF Staff
manifest
2e2ff2a verified
|
Raw History Blame Contribute Delete
11.8 kB
metadata
pretty_name: 'Supreme Court of the United States: Media Files Cited in Opinions'
license: unknown
language:
  - en
tags:
  - legal
  - law
  - supreme-court
  - united-states
  - government
  - court-documents
size_categories:
  - n<1K
configs:
  - config_name: media-files-cited-in-opinions
    data_files:
      - split: train
        path: media-files-cited-in-opinions/data/*.parquet

Supreme Court of the United States: Media Files Cited in Opinions

Every document file that the Supreme Court's website lists under Media Files Cited in Opinions, with the file itself, byte for byte, and its text. One row per file.

This is the media-files-cited-in-opinions config of incrediblecrab/us-scotus-opinions, which holds the collections the site's footer lists under Opinions, and this card is media-files-cited-in-opinions/README.md there.

Nothing here is edited by hand, and no text is corrected, normalized or generated. The pipeline and its tests are in github.com/incrediblecrab/us-dataset-products/tree/main/scotus, and this card is rendered from media-files-cited-in-opinions/manifest.json in the same commit.

Status

10 files (1.39 GB, 0 PDF pages); the listing of 2026-11-18T21:36:40Z linked 10 files from 10 entries. 10 rows are listed now and 0 are no longer listed.

Last complete run: 2026-11-18T21:36:40Z.

Text source Rows
no_text 10

Cross-check of text against pypdf: 0 files compared; 0 agree exactly on the sequence of non-whitespace characters, 0 agree after NFKD normalization of both sides, and the largest difference is 0 characters. pypdf reads the same text layer, so agreeing does not show that text matches the page, and differing does not show that text is wrong: see below.

What is verbatim

  • file is the file as the server sent it. file_sha256 lets you check it, and url says where it came from.
  • text is the PDF's own text layer as pdftotext -raw -enc UTF-8 (poppler) prints it, stored unchanged: line breaks, page headers and footers, and a form feed between pages. -raw keeps the characters in the order the file draws them. The default mode was rejected because it removes a hyphen that ends a line and joins the words; -layout was rejected because it dropped characters on a sample file. For an HTML page, text is the text the page renders.
  • The text layer is the publisher's, and text keeps it where it does not say what the page shows. A PDF can declare what a run of its glyphs says (ActualText), and pdftotext prints the declaration in place of the glyphs, so a glyph that a file declares to be no text is not in text. text is plain text and marks no type style, such as italics or underlining, and a letter set in small capitals is whatever character the file maps it to, so an opinion that prints JUSTICE ALITO in small capitals has "Justice Alito" in text. A code point from Unicode's Private Use Areas (U+E000 to U+F8FF, and planes 15 and 16) has no standard meaning: a file's font puts a ligature, a bullet, a dash or another sign there, and only file shows which. U+FFFD, the replacement character, says only that no character is known for a glyph. No row of text holds a Private Use Area code point or U+FFFD.
  • text holds only born-digital text. When one image covers at least half of a page that has text, the pipeline takes that text for an OCR layer made by whoever scanned the page, and the whole file's text goes to ocr_text, to be checked against file: on a scanned page it is OCR and not verbatim. The rule errs toward ocr_text, because it also catches a born-digital page that one image, such as a map or a photograph, covers at least half of. A file with no text on any page (no_text) has neither.
  • Every other text field (title, docket, entries) is the listing page's rendered text, with each run of ASCII whitespace collapsed to one space. No-break spaces, curly quotes and dashes are kept as the page has them.
  • term and date are read from those strings; the strings themselves stay in entries.

Use

from datasets import load_dataset
ds = load_dataset("incrediblecrab/us-scotus-opinions", "media-files-cited-in-opinions", split="train")
-- DuckDB, straight from the Hub, without the file bytes
SELECT id, term, date, docket, title, text_source FROM 'hf://datasets/incrediblecrab/us-scotus-opinions/media-files-cited-in-opinions/data/*.parquet' WHERE listed ORDER BY date DESC LIMIT 10;

Files

  • media-files-cited-in-opinions/data/{partition}.parquet: the rows, sorted by id. Partitions are by October Term.
  • media-files-cited-in-opinions/manifest.json: per partition, the row count, SHA-256 and summary counts; the files that failed and why; the listing pages read; the last 20 runs.

Schema

Column Type Description
id string The file's path on www.supremecourt.gov, percent-decoded and lower-cased, without the leading slash (the site ignores case). One row per file
partition string The Parquet file that holds the row (data/{partition}.parquet), assigned when the row was first stored and never changed
url string Where the file was fetched from
term int32 October Term, as the listing gives it (for U. S. Reports, the Term in the volume's listing entry; null where that entry names more than one); null where the listing gives none
date string ISO date read from the listing (decision, argument, order, filing or posting date, or the date a speech's title ends with); null where the listing gives no single date
docket string Docket number as the listing gives it; for Original Jurisdiction, the case number ("No. 1, Orig.")
title string Case name or document title as the listing gives it
entries large_string JSON list of every listing entry that links the file, in listing order: the listing page (listing), the link's href and text (link_text), its title attribute where the page gives one (link_title, the case summary on the opinions pages), its #page anchor (page), the table row it sits in (row, cells keyed by column header) or list item (item), the rendered line that holds it where that is neither (line), and the headings above it (section, group, subgroup). All strings are the page's text (rendered text, or the title attribute) with runs of ASCII whitespace collapsed to one space; no other character is changed. term, volume and year are numbers the pipeline read from the page or its address; docket (Argument Audio) is the link's text, and document_type (pdf, html, audio or video) is the kind of file the pipeline took the link for, in the collections whose listings mix them
listed bool Whether the latest complete listing still links the file. Rows are never deleted; a file the Court stops listing stays with listed = false
media_type string application/pdf, text/html for HTML documents, audio/mpeg for MP3 audio, or video/mp4 for MP4 video
file large_binary The file's bytes exactly as the server sent them
file_sha256 string SHA-256 of file
file_size int64 Length of file in bytes
etag string The ETag header the server sent with file
last_modified string The Last-Modified header the server sent with file
pages int32 Page count (pdfinfo); null for HTML, audio and video
image_pages int32 Pages on which one image covers at least half the page
ocr_pages int32 Pages that have text and are image pages, whose text the pipeline takes for an OCR layer
text_source string born_digital (the PDF's own text; in text), scanned or mixed (every page with text, or only some, is an image page, whose text the pipeline takes for OCR; in ocr_text), no_text (no page has text or the file is audio/video), or html
text large_string The file's own text, unchanged: the text layer of a born-digital PDF as pdftotext -raw prints it (form feed between pages), or the rendered text of an HTML page; see What is verbatim for where that differs from the page. Null when the pipeline takes the file's text for OCR
ocr_text large_string The same pdftotext -raw output when the pipeline takes the PDF's text for an OCR layer made by whoever scanned it (see ocr_pages). On a scanned page it is OCR, which misreads characters, so it is not verbatim. The same rule catches a born-digital page that one image, such as a map or a photograph, covers at least half of, and that page's text is the file's own
extractor string Program, version and options that produced text or ocr_text
xcheck_extractor string The independent extractor used for the cross-check (pypdf); null when there was no cross-check (only text from PDFs is cross-checked)
xcheck_equal bool Whether pypdf yields the same sequence of non-whitespace characters as text
xcheck_equal_nfkd bool The same comparison after NFKD normalization of both sides (ligatures and compatibility characters decomposed)
xcheck_delta int64 Characters (after NFKD, whitespace ignored) that one extraction has and the other lacks, counted without regard to order
metadata large_string JSON: the PDF's document information (pdfinfo), extraction notes, and the response headers kept
first_seen_at string When the pipeline first stored this file (UTC)
fetched_at string When the stored bytes were fetched (UTC)
checked_at string When the pipeline last confirmed the stored bytes are current: by fetching again or by the server's validators (UTC)
delisted_at string When the listing stopped linking the file (UTC); null while listed

Notes on this collection

  • The Court's page says: "On occasion, an opinion may cite to a media file, e.g., a video or audio file that is part of the record in the lower court. Those files are posted here." This collection stores only Court-hosted files linked by the page; it does not follow any outside source.
  • On September 27, 2026 the page linked 10 files: 9 MP4 videos and 1 MP3 recording. Audio and video rows store the bytes in file and have text_source = no_text; a response that does not open as MP3 audio or as an MP4 file (an ftyp box) is a failure, not a row.
  • Only one version of a file is kept: the one fetched at fetched_at. A run asks the server whether stored files changed only when it is run with --revalidate-all, which asks (HEAD) about every stored file and fetches again the ones whose ETag, Last-Modified or Content-Length moved, and reads every HTML page again. checked_at says when a row was last confirmed against the server, and the Hub's commit history holds earlier versions of the Parquet files.

How it is updated

The pipeline reads every listing page of the collection, fetches files it has not stored, updates the listing fields of files whose entries changed, and marks files that are no longer listed. Requests are paced to at least 1.1 seconds apart, as the site's robots.txt asks (Crawl-delay: 1), with the user agent us-dataset-products-scotus and a link to the repository.

License

The Court posts these files because an opinion cites them; its media page gives as an example "a video or audio file that is part of the record in the lower court". This dataset makes no claim about their copyright status. This card does not give legal advice.

The pipeline's code is under the MIT License (github.com/incrediblecrab/us-dataset-products/tree/main/scotus).