ResearchUser/CIVIC_culture

CIVIC-Culture Calibration Benchmark Dataset

language: English

task_categories:

• —evaluation
• —text-generation
• —text-classification
• — size_categories:
• —100-1k
• — tags:
• —culture
• —ethics
• —alignment
• —evaluation
• —responsible-ai
• —social-values ---

CIVIC-Culture Calibration Benchmark

Dataset Summary

The CIVIC-Culture Calibration Benchmark is a culturally grounded diagnostic dataset designed to evaluate how language models reason about normative social, ethical, and epistemic questions across cultures.

The dataset presents a set of culturally diagnostic prompts paired with region-specific normative completions, enabling systematic analysis of cultural alignment, value sensitivity, and cross-cultural reasoning behavior in large language models (LLMs).

It is intended for evaluation, calibration, and bias analysis, not for supervised training of factual knowledge.

Cultural Dimensions Covered

The dataset spans nine foundational cultural dimensions:

1. 1.Moral Reasoning
2. 2.Authority & Law
3. 3.Family Structure
4. 4.Truth & Justification
5. 5.Gender Roles
6. 6.Group vs. Individual
7. 7.Spirituality & Cosmology
8. 8.Education & Socialization
9. 9.Science & Epistemology

Each dimension contains multiple diagnostic prompts reflecting common normative tensions observed across societies.

Regional Perspectives

For each prompt, the dataset provides normative completions corresponding to four broad cultural regions:

• —West
• —Rest (Asia)
• —Middle East
• —Latin America (LATAM)

These responses reflect generalized cultural norms, not individual beliefs.

Dataset Structure

Data Format

The dataset is provided as a single CSV file: