PoSSUM Political Slant Dataset

Dataset author and curator: David Driscoll.
Version: improvement-20261018 · Publication: October 21, 2026
Project: PoSSUM

This is the 4,312-record dataset used for the PoSSUM research model: 2,199 passages derived from human MITweet annotations and 2,113 GPT-4-generated passages from Jyoti Shankar Nayak's Political Ideologies dataset. The source datasets, annotations and source texts retain their original authorship and attribution.

This release corresponds to the model's later human-text development revision, not the older 3,250-row PoSSUMDatasetv1.0 assembly. IBC, Manifesto and party-only proxy datasets are not included in this release.

Partitions

Split	MITweet	Synthetic	Total	Purpose
fit	1,224	956	2,180	Expert fitting
calibration	204	400	604	Probability calibration; PoSSUM used the human subset
gate	204	359	563	Mixture/router fitting
validation	408	398	806	Development model selection
topic_holdout	159	0	159	Frozen topic evaluation, already evaluated
Total	2,199	2,113	4,312

Labels are 0 = Left and 1 = Right. Preserve these roles; do not pool all rows for training and then claim the published holdout performance. The topic holdout was evaluated after model selection and is now exposed evaluation data, not a fresh unseen test for future development.

Sources and attribution

1. MITweet — Songtao Liu, Ziling Luo, Minghua Xu, Lixiao Wei, Ziyao Wei, Han Yu, Wei Xiang, and Bang Wang (2023). Ideology Takes Multiple Looks: A High-Quality Dataset for Multifaceted Ideology Detection. EMNLP, pp. 4200–4213. Paper and official citation, source repository. Source revision: ef6d2175ca562b2a8ec042a19d5a263374f2407b; CSV SHA-256: cbfaed2252479ae4bb231efe9a2012447619f767fd1d227d9ea9ec25986a539a.
2. GPT-4 Political Ideologies — Jyoti Shankar Nayak, political_ideologies. Dataset and upstream card, generation code. The upstream card describes GPT-4 generation and manual checks. Its publication year is not asserted here. The original upstream revision was not recorded; the preserved normalized local source snapshot has SHA-256 149ff799474036143b349805141667b925f3d8818c50e1e73952b39fe0f7468b.

Curation and labels

Passages contain 12–150 whitespace-delimited words. MITweet's related, noncentral annotations among facets I1, I3, I4, I5 and I6 must agree to produce a binary label. Upstream ideology 0 maps to Left; 2 maps to Right. These are derived facet labels, not newly collected direct overall U.S. left/right judgments. Synthetic upstream liberal=1 maps to PoSSUM Left=0, and conservative=0 maps to Right=1.

Automated semantic screening selected reasonably self-contained political stances and excluded unresolved context, conflicting selected facets and factual reporting without attributable ideological framing. This is model-assisted screening, not additional human annotation. Text normalization preserves the source wording; original labels and occurrence-level provenance are retained.

Exact and lexical near duplicates are grouped transitively before splitting. Synthetic examples keep historical roles. MITweet's Energy Crisis, Mass Shootings and Political Parties topics, plus linked duplicate components, form the topic holdout. Other human examples are partitioned 60/10/10/20 across fit/calibration/gate/validation. Topic themes can also occur in synthetic data; this is not a guarantee of wholly unseen policy concepts.

Files and schema

The root JSONL files offer a uniform schema for Hugging Face loading. original-splits.zip preserves the five original JSONL files and their manifest byte for byte. checksums.json records hashes of the original and normalized files before packaging; paths beginning with original/ refer to members of that archive, and paths beginning with data/ refer to the corresponding root JSONL filename. validation.json records packaging checks.

• id: original content identifier; unique across all five partitions.
• text, label, label_name, word_count: passage and binary target.
• source, source_url, label_kind: upstream origin and annotation type.
• topic, topics: source topics and retained occurrence topics.
• split: original PoSSUM role.
• group, group_ids, group_quality: original grouping metadata and its limitations.
• previously_used: historical flag recorded at assembly time; not a statement that the row remains unused today.
• provenance_json: lossless JSON serialization of the entire original record, including original labels, source identifiers, revisions, available curation decisions and occurrence provenance. Parse with json.loads.

from datasets import load_dataset
import json

ds = load_dataset("ddriscoll/PoSSUM")
fit = ds["fit"]
provenance = json.loads(fit[0]["provenance_json"])

Limitations and intended use

This is an English political-viewpoint research resource. Nearly half the rows are synthetic. Human data are strongly class- and topic-imbalanced: human validation has 366 Left and 42 Right passages and substantial abortion/women's-rights coverage. Automated selection can introduce errors and selection bias. Author/thread identifiers are unavailable, so the splits do not establish author-disjoint evaluation. Lexical duplicate screening does not certify removal of all semantic paraphrases.

Labels concern text, not a reliable personal political identity. Do not use these labels to profile the original authors. Tweets may contain political opinions, sensitive topics, names or offensive language; the source does not provide a complete anonymization guarantee. Binary probability does not measure ideological intensity, and a score near one-half does not establish neutrality. Broad U.S. political coverage and real-world generalization are not established.

Licensing

No blanket open license is asserted for this mixed-source collection. The upstream Political Ideologies card declares Apache License 2.0; its license text is included as LICENSE-APACHE-2.0.txt. PoSSUM changes include selection, whitespace normalization, label inversion and split assignment; upstream credits remain above.

MITweet's inspected repository does not provide an explicit dataset license. Its redistribution and downstream-use terms are therefore unresolved, and its presence here must not be interpreted as a new permission grant from its authors or original text creators. Source rights remain with their respective holders. The Hub metadata is license: unknown rather than incorrectly applying Apache 2.0 to MITweet. Check the original sources for applicable permissions and terms before reuse.

Citation

Cite this compilation and both upstream sources when using the full collection. The Nayak entry below is a descriptive dataset citation, not an upstream-provided BibTeX record.

@misc{driscoll2026possumdataset,
  author = {David Driscoll},
  title = {{PoSSUM} Political Slant Dataset},
  year = {2026},
  howpublished = {Hugging Face Datasets},
  url = {https://huggingface.co/datasets/ddriscoll/PoSSUM},
  note = {Version improvement-20261018}
}

@inproceedings{liu-etal-2023-ideology,
  author = {Liu, Songtao and Luo, Ziling and Xu, Minghua and Wei, Lixiao and Wei, Ziyao and Yu, Han and Xiang, Wei and Wang, Bang},
  title = {Ideology Takes Multiple Looks: A High-Quality Dataset for Multifaceted Ideology Detection},
  booktitle = {Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing},
  year = {2023},
  pages = {4200--4213},
  publisher = {Association for Computational Linguistics},
  doi = {10.18653/v1/2023.emnlp-main.256},
  url = {https://aclanthology.org/2023.emnlp-main.256/}
}

@misc{nayakpoliticalideologies,
  author = {Jyoti Shankar Nayak},
  title = {Political Ideologies: GPT-4-Generated Liberal and Conservative Statements},
  howpublished = {Hugging Face Datasets},
  url = {https://huggingface.co/datasets/JyotiNayak/political_ideologies},
  note = {Accessed October 21, 2026; original source revision not recorded in the PoSSUM snapshot}
}