World Values Survey Pilot
This benchmark adapts a compact subset of World Values Survey / WorldValuesBench ordinal questions to test cultural-value awareness in language models.
What It Measures
The WVS pilot asks each model to predict how a described survey respondent would answer value questions. The first shipped signal is context sensitivity: whether the model changes predictions across respondent contexts such as region, settlement type, and education.
The scoring code also supports Wasserstein-1 distance to human answer distributions. Raw WVS respondent data is not committed to this repo; reference distributions should be generated locally from official WVS data or a WorldValuesBench preparation step.
The repository persists a WVS source manifest listing official Wave 7 country questionnaire documents and download IDs. The active benchmark language files are derived pilot prompts; non-English files are Codex-translated and should not be described as official WVS translations.
Current Scope
The initial dataset contains 12 questions across social values, wellbeing, trust, and economic values in English, German, French, Spanish, Polish, Russian, Chinese, and Arabic. Each question is asked for four respondent contexts, for 48 model calls per run.
This is a pilot, not a replacement for the full WVS or WorldValuesBench. It is meant to establish the benchmark path in the existing runner before adding official translations and local reference distributions.