|
|
# Dataset Card for DEFSurveySim
|
|
|
|
|
|
## Dataset Summary
|
|
|
This dataset comprises carefully selected questions about human value preferences from major social surveys:
|
|
|
|
|
|
- **[World Values Survey (WVS-2023)](https://www.worldvaluessurvey.org/wvs.jsp)**: A global network of social scientists studying changing values and their impact on social and political life.
|
|
|
- **[General Social Survey (GSS-2022)](https://gss.norc.org/About-The-GSS)**: A survey of American adults monitoring trends in opinions, attitudes, and behaviors towards demographic, behavioral, and attitudinal questions, plus topics of special interest.
|
|
|
- **[Chinese General Social Survey (CGSS-2018)](http://cgss.ruc.edu.cn/English/Home.htm)**: The earliest nationwide and continuous academic survey in China collecting data at multiple levels of society, community, family, and individual.
|
|
|
- **[Ipsos Understanding Society survey](https://www.ipsos.com/en-uk/understanding-society)**: The preeminent online probability-based panel that accurately represents the adult population of the United States.
|
|
|
- **[American Trends Panel](https://www.pewresearch.org/our-methods/u-s-surveys/the-american-trends-panel/)**: A nationally representative online survey panel, consisting of over 10,000 randomly selected adults from across the United States.
|
|
|
- **[USA Today/Ipsos Poll](https://doi.org/10.25940/ROPER-31120147)**: Surveys a diverse group of 1,023 adults aged 18 or older, including 311 Democrats, 290 Republicans, and 312 independents.
|
|
|
- **[Chinese Social Survey](http://css.cssn.cn/css\_sy/)**: Longitudinal surveys focus on labor and employment, family and social life, and social attitudes.
|
|
|
|
|
|
The data supports research described in: [Towards Realistic Evaluation of Cultural Value Alignment in Large Language Models: Diversity Enhancement for Survey Response Simulation]()
|
|
|
|
|
|
## Purpose
|
|
|
This dataset enables:
|
|
|
1. Evaluation of LLMs' cultural value alignment through survey response simulation
|
|
|
2. Comparison of model-generated preference distributions against human reference data
|
|
|
3. Analysis of how model architecture and training choices impact value alignment
|
|
|
4. Cross-cultural comparison of value preferences between U.S. and Chinese populations
|
|
|
|
|
|
## Data Structure
|
|
|
```
|
|
|
DEF_survey_sim/
|
|
|
├── Characters/
|
|
|
│ ├── US_survey/
|
|
|
│ │ ├── Character.xlsx
|
|
|
│ │ └── ...
|
|
|
│ └── CN_survey/
|
|
|
│ ├── Character.xlsx
|
|
|
│ └── ...
|
|
|
├── Pref_distribution/
|
|
|
│ ├── usa_ref_score_all.csv
|
|
|
│ └── zh_ref_score_all.csv
|
|
|
├── Chinese_questionaires.txt
|
|
|
└── English_questionaires.txt
|
|
|
```
|
|
|
|
|
|
### Data Format Details
|
|
|
1. **Txt Files**: Original survey questions in txt format, maintaining survey integrity
|
|
|
2. **Characters**: Demographic breakdowns including:
|
|
|
- Age groups (under 29, 30-49, over 50)
|
|
|
- Gender (male, female)
|
|
|
- Dominant demographic characteristics per question
|
|
|
3. **Preference Distributions**: Statistical distributions of human responses for benchmark comparison
|
|
|
|
|
|
## Usage Guidelines
|
|
|
For implementation details and code examples, visit our [GitHub repository](https://github.com/alexc-l/DEF-Value-investigation).
|
|
|
|
|
|
## Limitations and Considerations
|
|
|
- Surveys were not originally designed for LLM evaluation
|
|
|
- Cultural context and temporal changes may affect interpretation
|
|
|
- Response patterns may vary across demographics and regions
|
|
|
- Limited construct validity when applied to artificial intelligence
|
|
|
|
|
|
## Contact Information
|
|
|
- Research inquiries: [email protected]
|
|
|
- Technical support: [GitHub Issues](https://github.com/alexc-l/DEF-Value-investigation/issues)
|
|
|
|
|
|
## Citation
|
|
|
```bibtex
|
|
|
TODO
|
|
|
``` |
