stanforddams/biglocal

Dataset Card for ICE Press Releases

A structured dataset of 965 press releases published by U.S. Immigration and Customs Enforcement (ICE), scraped from ice.gov. Created by Big Local News at Stanford University. This dataset gives journalists a machine-readable record of ICE's own public statements about enforcement actions.

Dataset Details

Dataset Description

ICE administrative data logs arrests, removals, and transfers but includes very little detail about individuals. Press releases are different: ICE routinely names people, lists ages, nationalities, and charges, and describes the specifics of each action in narrative form. This dataset captures that text in structured fields alongside extracted metadata, making it useful for named entity recognition, information extraction, and accountability journalism. The corpus is skewed toward 2025 and 2026 (956 of 965 records), reflecting the surge in ICE enforcement activity and press release volume during that period.

Dataset Sources

• Source: ICE Newsroom, Press Releases

Dataset Structure

Each row represents one ICE press release.

Field	Type	Coverage	Description
url	string	100%	Source URL on ice.gov
title	string	100%	Press release headline
subtitle	string	18%	Subtitle or secondary headline (sparse)
topics	string	99.9%	ICE topic category (see Topics section below)
date_raw	string	99.9%	Original date string as published
date_normalized	string	99.9%	Date normalized to MM/DD/YYYY
date_last_updated	string	100%	Date the page was last updated on ice.gov
location_full_text	string	98%	Full location string from the press release dateline
city	string	99.9%	City extracted from location
state	string	99.9%	State abbreviation extracted from location
full_text	string	99.9%	Full body text of the press release
image_urls	string	54%	URLs of images embedded in the press release
scraped_at	string	100%	ISO 8601 timestamp of when the page was scraped
blurb_list	string	0%	Not populated; reserved for future use
updated_date	string	0%	Not populated; reserved for future use

html.jsonl holds the raw source HTML for each press release and joins to the structured records above on url.

Information

Project Background

This dataset is Project 1 of two pitches submitted by Big Local News for the Datathon for Social Good 2026.

Deliverable:

• A structured dataset identifying the date, time, and location of each enforcement action described in a press release, along with as much identifying information as possible about individuals targeted e.g., name, age, sex, country of origin, etc.
• A data processing pipeline that can extract this type of data from press releases published in the future

File Guide

File	Description
index.csv	Primary structured dataset, 965 records, 15 fields
train.jsonl	965 raw HTML source pages scraped from ice.gov, one JSON object per line: {"url": ..., "html": ...}

Loading the Dataset

from datasets import load_dataset

ds = load_dataset("stanforddams/biglocal/")
html = load_dataset("stanforddams/biglocal/", "html")

Dataset Card Contact

Ryan Pitts, Managing Director, Big Local News ryanpitts@stanford.edu