Dataset Card for cultural-dimension-cover-letters

The cultural-dimension-cover-letters dataset contains cover letters modified to reflect different cultural dimensions based on Hofstede's framework. Created for evaluating implicit cultural preferences in large language models (LLMs) through job application assessment tasks.

Dataset Details

Dataset Description

This dataset features cover letters adapted to represent six cultural dimensions: Individualism/Collectivism, Power Distance, Masculinity/Femininity, Uncertainty Avoidance, Long-Term/Short-Term Orientation, and Indulgence/Restraint.

• Curated by: Ariba Khan, Stephen Casper, and Dylan Hadfield-Menell (MIT CSAIL)
• Language: English

Dataset Sources

• Base Cover Letters: ShashiVish/cover-letter-dataset
• Paper: Khan, A., Casper, S., & Hadfield-Menell, D. (2025). Randomness, Not Representation: The Unreliability of Evaluating Cultural Alignment in LLMs. arXiv:2503.08688. https://arxiv.org/abs/2503.08688

Uses

Direct Use

For evaluating cultural biases in LLMs, studying cultural effects on AI evaluation, and researching evaluation context effects on AI preferences.

Dataset Structure

Contains 600 comparison pairs (100 per dimension), with each pair featuring two variations of the same cover letter representing opposite cultural dimensions:

1. Individualism vs. Collectivism (idv.xlsx)
2. Long-Term vs. Short-Term Orientation (lto.xlsx)
3. Masculinity vs. Femininity (mas.xlsx)
4. High vs. Low Power Distance (pdi.xlsx)
5. High vs. Low Uncertainty Avoidance (uai.xlsx)
6. Indulgence vs. Restraint (ivr.xlsx)

Dataset Creation

Source Data

Data Collection and Processing

The dataset was created using the ShashiVish/cover-letter-dataset, which comprises 813 cover letters predominantly from technology sector job applications. Using GPT-4o, we generated culturally distinct versions of each cover letter across Hofstede's six cultural dimensions.

The adaptation process used a standardized prompt: "You are a professional cover letter writer. Rewrite the cover letter with a [dimension] tone. Maintain the same content and length of writing as the original cover letter."

The process generated 12 variations per cover letter (two poles for each cultural dimension), randomly selected 100 letters per dimension yielding 600 comparison pairs, and manually verified a subset to ensure they accurately reflected the intended cultural dimensions.

Citation

@article{khan2025randomness, title={Randomness, Not Representation: The Unreliability of Evaluating Cultural Alignment in LLMs}, author={Khan, Ariba and Casper, Stephen and Hadfield-Menell, Dylan}, journal={arXiv preprint arXiv:2503.08688}, year={2025}, url={https://arxiv.org/abs/2503.08688} }

Dataset Card Contact

Ariba Khan (akhan02@mit.edu), Stephen Casper (scasper@mit.edu)