README.md

metadata

pretty_name: Verasight U.S. Public Opinion Data Library
license: cc-by-4.0
language:
  - en
multilinguality:
  - monolingual
source_datasets:
  - original
size_categories:
  - 10K<n<100K
tags:
  - tabular
  - text
  - datasets
  - pandas
  - dask
  - mlcroissant
  - public-opinion
  - polling
  - polls
  - survey
  - survey-research
  - social-science
  - united-states
  - artificial-intelligence
  - politics
  - health
  - economics
  - culture
configs:
  - config_name: questions
    default: true
    data_files:
      - split: train
        path: data/questions.parquet
  - config_name: estimates
    data_files:
      - split: train
        path: data/estimates.parquet
  - config_name: pages
    data_files:
      - split: train
        path: data/pages.parquet

A source-linked index of what U.S. adults think

The Verasight Data Library makes original U.S. public opinion research searchable and ready for analysis. Discover questions and weighted toplines across AI & Tech, Culture, Health, Money, Politics, Sports, then follow every record to a human-readable finding and its verified primary source report.

Explore findings, search topics, and cite the research at data.verasight.io

Coverage at a glance

Survey waves	Years	Published questions	Topline estimates	Data Library pages	Categories
37	2022 to 2026	1,697	8,695	1,430	6

Browse all questions · Search the library · Read the methodology

Explore by topic: AI & Tech · Culture · Health · Money & Finance · Politics · Sports

What you can do with it

If you are...	Use the dataset to...
A journalist or editor	Find citable public opinion numbers and trace them to exact wording and verified sources.
A researcher or social scientist	Filter questions by topic, field dates, sample size, and weighted topline estimates.
An AI, search, or RAG builder	Index current, sourced opinions while retaining URLs for retrieval and attribution.
A policy, communications, or insights team	Explore how U.S. adults respond across six recurring topic areas.

Why this dataset

• Original measured opinion: findings come from Verasight survey research, not synthetic responses or a collection of prompts without respondent data.
• Question-level provenance: question wording, field dates, base sizes, source status, verified report links, and Data Library links travel with the records.
• Human and machine-readable: Parquet tables support analysis, while the linked Data Library pages add summaries, context, and citation paths.
• Transparent methodology: Verasight publishes its recruitment, weighting, margin-of-error, wording, and topline practices and participates in the AAPOR Transparency Initiative.
• Audited public boundary: this package contains only reviewed aggregate findings. It excludes respondent-level records, raw survey files, unpublished questions, and free-text responses.

Use the right layer

Layer	Best for	Role
This Hugging Face dataset	Filtering, analysis, indexing, and reproducible discovery	Machine-readable distribution layer
Verasight Data Library	Reading, browsing, sharing, and citing findings	Canonical interpretation layer
Verified Verasight source reports	Exact questionnaires, toplines, and survey methodology	Primary source record

Dataset contents

• questions: 1,697 published source questions with wording, field dates, methodology, source citations, and Data Library URLs.
• estimates: 8,695 weighted topline option estimates, each with a direct Data Library URL.
• pages: 1,430 published featured-topic and standalone question pages with summaries and canonical URLs.

All three subsets use the train split only as a Hugging Face loading convention. The data is not divided into machine-learning training and test sets.

Load with Datasets

from datasets import load_dataset

questions = load_dataset("Verasight/verasight-data-library", "questions", split="train")
estimates = load_dataset("Verasight/verasight-data-library", "estimates", split="train")
pages = load_dataset("Verasight/verasight-data-library", "pages", split="train")

Find matching questions and continue to the human-readable finding:

for row in questions:
    if "artificial intelligence" in row["title"].lower():
        print(row["title"])
        print(row["data_library_url"])

Key fields

questions

• title and context: published wording and supporting prompt text.
• field_date_start, field_date_end, base_n, and margin_of_error: survey context needed for interpretation.
• canonical_url: the verified source report or question anchor.
• data_library_url: the preferred human-readable page for this question.
• page_urls: every Data Library page that uses the question.
• categories: the topic areas assigned by the Data Library.

estimates

• option_label: the published response option.
• weighted_share: the estimate on a 0 to 1 scale.
• weighted_percent: the same estimate on a 0 to 100 scale.
• unweighted_n: the available unweighted count for that option.
• canonical_url and data_library_url: source and interpretation paths.

pages

• title, dek, and summary: the published editorial framing.
• url and markdown_url: canonical HTML and machine-readable page versions.
• source_question_ids and source_wave_ids: provenance for the page.

Source and scope

The release boundary is the site build's structured-output.json manifest. That manifest contains the reviewed public surface, not the broader canonical or confidential pipeline data. Source reports remain the canonical survey references. Each applicable record includes a public report URL.

Every question and estimate includes both a Data Library URL and a verified canonical source report URL. The public release exporter rejects pending or unavailable source records before creating the package.

Interpretation and methodology

weighted_share is a proportion from 0 to 1. weighted_percent is the same estimate on a 0 to 100 scale. unweighted_n is the available unweighted option count. Consult the linked source and methodology before comparing estimates or drawing conclusions.

For methodology details, including panel recruitment, respondent verification, weighting, margins of error, and transparency practices, see the Data Library methodology. Methods and question context can vary by survey, so use field dates, base sizes, and linked source reports when making comparisons.

Citation and attribution

For a specific finding, cite the record's data_library_url and consult its canonical_url source report. This preserves the wording, timing, methodology, and interpretation that belong with the estimate.

Recommended dataset attribution: Verasight, Verasight U.S. Public Opinion Data Library, data.verasight.io and Verasight/verasight-data-library, licensed under CC BY 4.0. Indicate if you made changes.

Limitations and responsible use

• The tables contain aggregate toplines, not respondent-level microdata.
• Published percentages describe measured opinion during the listed field dates. They do not establish causation or predict future opinion.
• The dataset does not include every question Verasight has fielded. It includes only findings approved for the public Data Library.
• The train split is a loading convention, not a recommendation to use the records as model-training data.

Release freshness

This release covers 37 source waves from 2022 to 2026. Source build timestamp: 2026-10-23T18:34:54.108Z. The clean URLs stored in the tables are stable citation targets even when this snapshot is superseded by a later release.

License

Unless otherwise noted, the dataset files and original documentation in this repository are licensed under CC BY 4.0. See LICENSE for scope, attribution, and exclusions.