The surveys SynthBench scores against
The human survey datasets behind every score, with the source, coverage, demographics, and redistribution policy for each.
Registered datasets
Nine datasets are registered. Four are scored and make up the SynthBench Index (In the Index). NTIA is scored for reference only: its questions are yes/no facts about internet use, not opinions (Reference only). The other four have adapters but no published runs yet (Not scored yet).
Each card also shows the dataset's redistribution tier. full datasets ship per-question distributions publicly. gated datasets require a research-use sign-in for per-question data.
OpinionsQA
In the Index gatedThe primary ground-truth dataset. Multiple-choice questions covering political, social, and economic attitudes from nationally representative US surveys.
- Questions
- 1,498 (200 core)
- Source
- Pew American Trends Panel
- Scope
- US population
- Coverage
- 2017–2022 (15 survey waves)
- Demographics
- 56 groups across 11 attributes
GlobalOpinionQA
In the Index gatedCross-national attitudes from Pew Global Attitudes surveys. Tests whether models capture how opinion differs between countries, beyond US demographics.
- Questions
- 2,556
- Source
- Pew Global Attitudes
- Scope
- 138 countries
- Coverage
- Cross-national
- Demographics
- Country-level cultural variation
SubPOP
In the Index gatedExtended dataset focused on demographic subgroup variation. Tests whether models can differentiate between fine-grained population segments.
- Questions
- 3,362
- Source
- 22 US subpopulations
- Scope
- US demographic subgroups
- Coverage
- Cross-sectional
- Demographics
- 11 attributes: age, education, gender, ideology, party, income, religion, attendance, region, marital status, citizenship
GSS
In the Index fullGeneral Social Survey cumulative microdata from NORC. It is the only public-domain adapter that ships per-respondent rows, which allows convergence curves from real sampling alongside the bootstrap baseline.
- Source
- NORC General Social Survey
- Scope
- US population
- Coverage
- 1972–present (cumulative file)
- Demographics
- Individual-level microdata (microdata-capable)
NTIA Internet Use Supplement
Reference only fullNTIA Computer & Internet Use Supplement to the Current Population Survey. U.S. Government public-domain work under 17 USC 105; per-question distributions ship openly.
- Source
- US Census / NTIA CPS Supplement
- Scope
- US households
- Coverage
- Biennial waves, 1994–present
- Demographics
- Standard CPS demographics
WVS (World Values Survey)
Not scored yet gatedWorld Values Survey Wave 7. Cross-cultural value orientations across national populations. Research-use license. Per-question artifacts are gated behind JWT auth.
- Source
- World Values Survey Association
- Scope
- Cross-national (Wave 7, 2017–2022)
- Coverage
- Wave 7 (2017–2022)
- Demographics
- Country + standard demographic cells
Eurobarometer
Not scored yet gatedEurobarometer Consumer Modules. EU-wide consumer conditions, digital services, and sustainability attitudes. GESIS restricts redistribution, so per-question data is served from a gated origin.
- Source
- European Commission / GESIS
- Scope
- EU member states
- Coverage
- Quarterly (Standard / Flash / Special)
- Demographics
- Country + EU-harmonized demographic cells
Michigan (Surveys of Consumers)
Not scored yet gatedUniversity of Michigan Surveys of Consumers: consumer sentiment, inflation expectations, and household finance. Academic use terms. Per-question data is gated.
- Source
- University of Michigan SRC
- Scope
- US consumers
- Coverage
- Monthly cadence
- Demographics
- Standard consumer-survey demographics
Pew Technology
Not scored yet gatedPew Research Internet & Technology waves: tech adoption, digital privacy, social media, and AI attitudes. Pew redistribution terms. Per-question data is gated.
- Source
- Pew Research Center Internet & Technology
- Scope
- US population (American Trends Panel)
- Coverage
- 10 tech waves (W86–W113)
- Demographics
- ATP demographic weights
Dataset redistribution policy
Each upstream dataset has its own license terms. SynthBench applies them at publish time using four tiers:
- full: per-question distributions are published openly (public-domain datasets).
- gated: per-question distributions are served from a Cloudflare Worker that validates a Supabase JWT, so research-use sign-in is required. No distribution is served anonymously.
- aggregates_only: aggregate scores are public, and no per-question artifact ships at all. Gating is not a substitute for missing redistribution permission.
- citation_only: question metadata only. Every downstream metric is suppressed.
Aggregate metrics (SPS, JSD, ranks, cross-model parity) are derived quantities and remain public regardless of tier.
| Dataset | Tier | Source & license |
|---|---|---|
| eurobarometer | gated | European Commission / GESIS — Eurobarometer Consumer Modules (license) |
| globalopinionqa | gated | Durmus et al. 2023, Anthropic — llm_global_opinions (license) |
| gss | full | NORC at the University of Chicago — General Social Survey (license) |
| michigan | gated | Survey Research Center, University of Michigan — Surveys of Consumers (license) |
| ntia | full | NTIA Internet Use Survey (U.S. Government work, 17 USC 105) (license) |
| opinionsqa | gated | Santurkar et al., ICML 2023 — Whose Opinions Do LLMs Reflect? (derived from Pew American Trends Panel) (license) |
| pewtech | gated | Pew Research Center — Internet & Technology (American Trends Panel) (license) |
| subpop | gated | Suh et al., ACL 2025 — SubPOP: Subpopulation-Level Opinion Prediction (license) |
| wvs | gated | World Values Survey Association — WVS Wave 7 (2017-2022) (license) |
The methodology page explains how these datasets are scored: metrics, baselines, the private holdout split, and anti-gaming defences.