SpecNorm-Bench

A matched-pair benchmark for diagnosing subgroup cultural norm awareness in LLMs.

Each item pairs a GEN (generic national persona) and SPEC (specific subgroup persona) version of the same situation and question, sourced and derived from the NCLB/CultureAtlas dataset. See pipeline_design.md (project workspace) for full 9-stage construction details.

• Rows: 3,919 matched pairs (evaluable set — items with a resolved label_generic of yes/no)
• Countries: 146
• Cohorts: convergent (label_specific == label_generic, n=2,971), divergent (label_specific != label_generic, n=948)
• Note: an additional 609 items where label_generic == "uncertain" (the "neutral cohort") were produced by the construction pipeline but are excluded from this release since they aren't used in any analysis.

Columns

Column	Description
country, subcountry_region_extraction, ethnicity_extraction, religion_extraction, age_extraction, gender_extraction, marital_status_extraction, occupation_extraction	Demographic metadata extracted from the source NCLB statement
category, topic_norm, interactive_nature_categorization, topic_categorization	Topic/category annotations
statement, sentence	Original NCLB cultural norm statement
specificity_category, n_dims	Specificity level (Low/Medium/High) and number of non-null demographic dimensions
situation, question	Generated, persona-neutral scenario and yes/no question shared by both GEN and SPEC
persona_specific, persona_generic	The two persona variants
label_specific, label_generic	Ground-truth labels (yes/no/uncertain)
label_reasoning, evidence	LLM label rationale and retrieved web evidence (for label_generic)
url, vector, orig_idx	Source provenance / internal indexing fields

Known limitations (disclose before citing)

• label_specific is inherited directly from the original NCLB/CultureAtlas source annotation and was not independently re-verified against the LLM-generated situation text. A small fraction of items (~0.2% in a keyword audit) depict the persona acting in a different region/country's normative context than their own subgroup (e.g. a persona traveling abroad and following local host-culture norms), which can make the inherited label a weaker match for the generated situation.
• Human validation (Stage 5/9 in the construction pipeline) was performed on stratified subsamples only (500 and 300 items respectively), not the full set.

This is a research-in-progress artifact accompanying an unpublished manuscript; the repo is private pending completion of the paper.