Dataset Card for TwinViews-13k

This dataset contains 13,855 pairs of left-leaning and right-leaning political statements matched by topic. The dataset was generated using GPT-3.5 Turbo and has been audited to ensure quality and ideological balance. It is designed to facilitate the study of political bias in reward models and language models, with a focus on the relationship between truthfulness and political views.

Dataset Details

Dataset Description

TwinViews-13k is a dataset of 13,855 pairs of left-leaning and right-leaning political statements, each pair matched by topic. It was created to study political bias in reward and language models, with a focus on understanding the interaction between model alignment to truthfulness and the emergence of political bias. The dataset was generated using GPT-3.5 Turbo, with extensive auditing to ensure ideological balance and topical relevance.

This dataset can be used for various tasks related to political bias, natural language processing, and model alignment, particularly in studies examining how political orientation impacts model outputs.

• Curated by: Suyash Fulay, William Brannon, Shrestha Mohanty, Cassandra Overney, Elinor Poole-Dayan, Deb Roy, Jad Kabbara
• Language(s) (NLP): en
• License: cc-by-4.0

Dataset Sources

• Repository: https://github.com/sfulay/truth_politics
• Paper: https://arxiv.org/abs/2409.05283

Uses

Direct Use

This dataset is suitable for:

• Studying political bias in reward models and large language models (LLMs).
• Evaluating alignment techniques for LLMs, especially regarding truthfulness and political bias.
• Training and/or evaluating models in the context of political discourse analysis.
• Research on how political views and alignment objectives interact in AI systems.

Out-of-Scope Use

This dataset is not suitable for tasks requiring very fine-grained or human-labeled annotations of political affiliation beyond the machine-generated left/right splits. Notions of "left" and "right" may also vary between countries and over time, and users of the data should check that it captures the ideological dimensions of interest.

Dataset Structure

The dataset contains 13,855 pairs of left-leaning and right-leaning political statements. Each pair is matched by topic, with statements generated to be similar in style and length. The dataset consists of the following fields:

• l: A left-leaning political statement.
• r: A right-leaning political statement.
• topic: The general topic of the pair (e.g., taxes, climate, education).

Dataset Creation

Curation Rationale

The dataset was created to fill the gap in large-scale, topically matched political statement pairs for studying bias in LLMs. It allows for comparison of how models treat left-leaning versus right-leaning perspectives, particularly in the context of truthfulness and political bias.

Source Data

Data Collection and Processing

The data was generated using GPT-3.5 Turbo. A carefully designed prompt was used to generate statement pairs that were ideologically representative of left-leaning and right-leaning viewpoints. The statements were then audited to ensure relevance, ideological alignment, and quality. Topic matching was done to ensure the statements are comparable across the political spectrum.

In summary:

• Generated using GPT-3.5 Turbo.
• Audited for ideological and topical relevance.
• Final dataset filtered and structured to ensure left/right statement parity.

Who are the source data producers?

The dataset was generated by GPT-3.5 Turbo, with extensive auditing performed by the dataset creators at MIT.

Personal and Sensitive Information

The dataset consists of machine-generated political statements and does not contain any personal or sensitive information.

Bias, Risks, and Limitations

Users of the dataset should be aware of certain limitations:

• Source context: Notions of what is political and the left/right ideological spectrum are context-specific and vary between countries and over time. Our dataset and its notions of politics and ideology come from the US in the early 2020s and may not generalize to other cultures or other time periods.
• Generated content: Since the statements were generated by GPT-3.5 Turbo, they may not fully capture the nuance or complexity of real-world political discourse. It is also possible that the dataset may contain stylistic or lexical artifacts correlated with political bias, though our evaluation has not identified any such artifacts.

Citation

BibTeX:

@inproceedings{fulayRelationshipTruthPolitical2024,
  author       = {Fulay, Suyash and Brannon, William and Mohanty, Shrestha and Overney, Cassandra and Poole-Dayan, Elinor and Roy, Deb and Kabbara, Jad},
  title        = {On the Relationship between Truth and Political Bias in Language Models},
  booktitle    = {Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP '24)},
  year         = {2024},
  month        = nov,
  publisher    = {Association for Computational Linguistics},
  note         = {arXiv:2409.05283},
  abstract     = {Language model alignment research often attempts to ensure that models are not only helpful and harmless, but also truthful and unbiased. However, optimizing these objectives simultaneously can obscure how improving one aspect might impact the others. In this work, we focus on analyzing the relationship between two concepts essential in both language model alignment and political science: \textit{truthfulness} and \textit{political bias}. We train reward models on various popular truthfulness datasets and subsequently evaluate their political bias. Our findings reveal that optimizing reward models for truthfulness on these datasets tends to result in a left-leaning political bias. We also find that existing open-source reward models (i.e. those trained on standard human preference datasets) already show a similar bias and that the bias is larger for larger models. These results raise important questions about both the datasets used to represent truthfulness and what language models capture about the relationship between truth and politics.}
}

APA:

Fulay, S., Brannon, W., Mohanty, S., Overney, C., Poole-Dayan, E., Roy, D., & Kabbara, J. (2024). On the Relationship between Truth and Political Bias in Language Models. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP '24). Association for Computational Linguistics.

Glossary

• Left-leaning: Political statements generally associated with progressive or liberal views.
• Right-leaning: Political statements generally associated with conservative or traditional views.
• Political Bias: A model's tendency to favor one political ideology over another in its outputs.

Dataset Card Authors

William Brannon, wbrannon@mit.edu

Dataset Card Contact

• William Brannon, wbrannon@mit.edu
• Suyash Fulay, sfulay@mit.edu