WestlakeNLP/PAPI-300K

Personality Alignment with Personality Inventories (PAPI) Dataset

Overview

The Personality Alignment with Personality Inventories (PAPI) dataset is a large-scale, real-world collection of personality assessments designed to evaluate the consistency between Personality-aligned Language Models (LLMs) and human preferences. This dataset contains 307,313 samples of multiple-choice questionnaire responses, providing a comprehensive resource for research in personality psychology and AI alignment.

Dataset Description

Size and Structure

• Total samples: 307,313 (Dev-Set and Test-Set)
• Each sample contains:
  • IPIP-NEO-120 questionnaire responses
  • IPIP-NEO-300 questionnaire responses

Question Format

• Each question is associated with one of the five major personality dimensions
• Questions correlate positively or negatively with specific behavioral patterns or preferences
• Five response options ranging from "Very Accurate" to "Very Inaccurate"

Demographic Information

• Gender distribution:
  • Female: ~60% (185,149 participants)
  • Male: ~40% (122,164 participants)
• Age range: 10 to 99 years
• Average age: 25.19 years
• Geographical diversity: Participants from various countries including the USA, UK, France, India, and China

Data Collection

• Source: International Personality Item Pool (IPIP)
• Collection period: 1998 to 2019
• Method: Online surveys
• Survey duration: 30 to 50 minutes per participant
• Anonymity: All responses are anonymous

Data Processing

• K-Means clustering applied to create a representative Test-Set
• 300 representative clusters selected as Test-Set
• Remaining data released as Dev-Set

Evaluation Methods

• Behavioral difference score used as an evaluation metric
• Scoring function assigns values from 5 to 1 for responses ranging from "Very Accurate" to "Very Inaccurate"
• Aligned Score calculated for each Big Five trait

Ethical Considerations

• Voluntary participation with no monetary compensation
• Data anonymization: No personally identifiable information (PII) included
• Compliance with data protection regulations (e.g., GDPR)
• Adherence to ethical principles outlined in the Belmont Report

Dataset Analysis

Detailed analysis of the dataset is provided, including:

• Demographic distributions (age, gender, country)
• Personality trait distributions for the Big Five traits
• Comparison between the Total dataset and the Test-Set

Citation

If you use this dataset in your research, please cite:

@misc{zhu2024personalityalignmentlargelanguage,
      title={Personality Alignment of Large Language Models}, 
      author={Minjun Zhu and Linyi Yang and Yue Zhang},
      year={2024},
      eprint={2408.11779},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2408.11779}, 
}

Contact

For questions or additional information about the PAPI dataset, please contact:

zhuminjun@westlake.edu.cn

Acknowledgments

We thank the administrators of the International Personality Item Pool (IPIP) for granting permission to use their items, scales, and inventories. We also express our gratitude to all the volunteers who participated in this study, contributing to the advancement of personality research and AI alignment.