README.md

---
license: cc-by-nc-sa-4.0
extra_gated_prompt: >-
  SubPOP dataset is a subpopulation-level opinion distribution dataset derived
  from a limited portion of public opinion survey questions from Pew Research
  Center's American Trends Panel and NORC at the University of Chicago's General
  Social Survey (2022 cross-sectional survey).

   By requesting access to this repository, you explicitly and clearly agree to adhere to the following terms:
   1. Acknowledge the source of the data with express reference to the owner of the data, in accordance with the following citation: 'Pew Research Center’s American Trends Panel', 'Davern, Michael; Bautista, Rene; Freese, Jeremy; Herd, Pamela; and Morgan, Stephen L. General Social Survey 1972-2022. [Machine-readable data file]. Principal Investigator, Michael Davern; Co-Principal Investigators, Rene Bautista, Jeremy Freese, Pamela Herd, and Stephen L. Morgan. NORC ed. Chicago, 2024. 1 datafile and 1 codebook (2022 Release 3a).'
   2. Use the dataset solely for (1) research, scholarly or academic purposes, or (2) user's own personal, non-commercial use.
   3. Do not use the Data in any manner that implies, suggests, or could otherwise be perceived as attributing a particular policy or lobbying objective or opinion to the survey research centers.
   4. Include the following disclaimer: “The opinions expressed herein, including any implications for policy, are those of the author and not of the survey research centers.”
   5. Acknowledges that, as between the parties, the survey research centers are the sole and exclusive owners of all right, title and interest in data. Except for the limited license granted herein, this agreement does not give use any right, title or interest in the data.
   6. Accurately and fully complete the required fields for access requests, including affiliation, country, and intended purpose.
   7. Agree to not use the dataset for any experiments or research activities that could cause harm to human subjects.
extra_gated_fields:
  Affiliation: text
  Country: country
  Intended purpose (select an option):
    type: select
    options:
    - Research
    - Education
    - label: Other
      value: other
  Intended purpose (describe in text): text
  I agree to use this dataset for non-commercial and scholarly or personal use ONLY: checkbox
language:
- en
size_categories:
- 10K<n<100K
formats:
- csv
---

## Dataset Card for SubPOP

The dataset is from the paper, [Language Model Fine-Tuning on Scaled Survey Data for Predicting Distributions of Public Opinions](https://arxiv.org/abs/2502.16761) by Joseph Suh, Erfan Jahanparast, Suhong Moon, Minwoo Kang, and Serina Chang.

### Dataset Summary

The dataset contains a subset of survey questions about social issues and opinions in the United States adapted from [American Trends Panel](https://www.pewresearch.org/the-american-trends-panel/) and [General Social Survey](https://gss.norc.org/us/en/gss/get-the-data.html).

It consists of two parts, *SubPOP-Train* and *SubPOP-Eval*, which are derived from American Trends Panel and General Social Survey, respectively.

**Questions**. We include 3,229 questions for *SubPOP-Train*, and 133 questions for *SubPOP-Eval*.

**Subpopulations**. Following [OpinionQA dataset](https://github.com/tatsu-lab/opinions_qa), we include 22 subpopulations across 8 attributes: region, income, education, race/ethnicity, gender, religion, political party affiliation, and political ideology.

In our paper, we use this dataset to evaluate the large language models' capability to predict the subpopulation-level opinion response distributions.

### File Structure

Each list element in the files `subpop_train.jsonl` and `subpop_eval.jsonl` is a dictionary corresponding to a subpopulation's response distribution to a survey question. The dictionary items are:

- **qkey** (`str`): the question's ID, adopted from the original survey data.

- **attribute** (`str`): demographic or ideology variable defining the subpopulation, such as `CREGION` (region), `POLPARTY` (political party affiliation).

- **group** (`str`): subpopulation, such as `Northeast` (population living in the U.S. Northeast), `Democrat` (population identified as Democrat).

- **question** (`str`): the survey question string.

- **options** (`List[str]`): the list of option choices presented at the time of survey.

- **ordinal** (`List[float]`): ordinal structure of the option choices. Following [OpinionQA](https://github.com/tatsu-lab/opinions_qa), we include the ordinal structure for *SubPOP-Eval*. For example, a list of options `Strongly agree`, `Agree`, `Neutral`, `Disagree`, `Strongly disagree` is mapped to floating point number 1.0 through 5.0.

- **responses** (`List[float]`): the response distribution of the group to the question. It is a post-stratification weighted distribution, and values in the list are normalized to a sum of 1.0. For more information about weighting, please refer to [For Weighting Online Opt-in Samples, What Matters Most?](https://www.pewresearch.org/methods/2018/01/26/how-different-weighting-methods-work/).

- **refusal_rate** (`float`): the fraction of respondents who selected 'Refused to answer' choice at the time of survey.

The dataset is further described in our [paper](https://arxiv.org/abs/2502.16761) and the [GitHub repository](https://github.com/josephJeesungSuh/subpop/).

### Citation

```
@article{suh2025language,
  title={Language Model Fine-Tuning on Scaled Survey Data for Predicting Distributions of Public Opinions},
  author={Suh, Joseph and Jahanparast, Erfan and Moon, Suhong and Kang, Minwoo and Chang, Serina},
  journal={arXiv preprint arXiv:2502.16761},
  year={2025}
}
```

### Contact

- josephsuh@berkeley.edu