Populism-LLM

A cross-validated LLM annotation of European party manifestos on populism and liberalism dimensions. Three independent large language models — Claude Sonnet 4.6 (Anthropic), GPT-4.1-mini (OpenAI), and Gemini Flash (Google) — score every CMP manifesto using an identical strict JSON schema. Each score is backed by a verbatim quote and contextual snippet.

🚧 Preview / draft release. A peer-reviewed paper-companion v1.0 release with DOI is forthcoming. Until then, contact sebastian.stoeckl@uni.li before citing.

Scope

• 3,327 manifestos from the Manifesto Project Database (CMP / WZB), 1945–2025
• 50+ countries, ~15 languages
• 12 dimensions scored per chunk: anti-elitism, people-centrism, manichaean worldview, populism (overall), ideology (left / right / centrist / overall), political / social / economic / financial-market liberalism, liberalism (overall)
• Per-chunk evidence: every non-null score is backed by a 3–8 word verbatim quote and a ~3-sentence context snippet

Files

Start here — the analysis-ready table:

File	Rows	Contents
data/manifesto_scores.parquet	3,327 manifestos	One row per manifesto. CMP metadata + populism/liberalism scores (the paper's baseline) + per-model columns. This reproduces the paper's headline regressions.
data/manifesto_scores.csv	3,327 manifestos	Identical, for non-R/Python users.

Raw per-chunk annotations (to rebuild the aggregate yourself):

File	Rows	Model	Family
data/raw_sonnet.parquet	6,561 chunks → 3,327 manifestos	Claude Sonnet 4.6	Anthropic
data/raw_gpt41mini.parquet	6,561 chunks → 3,327 manifestos	GPT-4.1-mini	OpenAI
data/raw_gemini.parquet	6,561 chunks → 3,327 manifestos	Gemini Flash (latest)	Google

manifesto_scores uses the paper's headline aggregation: a confidence-weighted mean over each model's verified chunks (w = pmax(populism_confidence, 0.1)), then a plain mean across the three models. The raw_* files let you reproduce that aggregation, inspect the verbatim evidence, or apply a different aggregation rule.

Each raw_* row is one manifesto chunk. Manifestos longer than 20k tokens are split into multiple chunks with 2k-token overlap; the row's chunk_id identifies the chunk and doc_id_root ties chunks back to a manifesto.

Schema (per chunk × dimension)

For each of the 12 dimensions {d}, the file contains:

Column	Type	Notes
{d}_score	int 0–10, nullable	null when evidence_status = "no_evidence"
{d}_status	enum	verified, mixed, or no_evidence
{d}_quote	string	Verbatim 3–8-word quote from the chunk text
{d}_context	string	~3-sentence surrounding context (max 80 words)

Plus per-chunk confidences (populism_confidence, pop_ideology_confidence, liberalism_confidence) and free-text justifications.

Loading

The baseline measure — one line, no aggregation needed:

library(arrow)
scores <- read_parquet(
  "https://huggingface.co/datasets/sstoeckl/populism-llm/resolve/main/data/manifesto_scores.parquet"
)
# scores$populism_overall, scores$liberalism_overall, scores$lib_economic, ...
# grouped by scores$countryname / scores$party / scores$year

import pandas as pd
scores = pd.read_parquet(
  "https://huggingface.co/datasets/sstoeckl/populism-llm/resolve/main/data/manifesto_scores.parquet"
)

The raw per-chunk annotations (with verbatim evidence):

df <- read_parquet(
  "https://huggingface.co/datasets/sstoeckl/populism-llm/resolve/main/data/raw_sonnet.parquet"
)

manifesto_scores columns

• Metadata: doc_id, countryname, partyname, partyabbrev, party, parfam, rile, date, year, token_count
• Baseline (cross-model mean, 0–10): populism_overall, liberalism_overall, the populism sub-dims (pop_anti_elitism, pop_people_centrism, pop_manichean), ideology (pop_ideology_overall and left/right/centrist), and the liberalism sub-dims (lib_political, lib_social, lib_economic, lib_financial_market)
• Per-model: every dimension above also as <dim>_sonnet, <dim>_gpt41mini, <dim>_gemini

Methodology

Pipeline

1. Chunking — ≤20k tokens per chunk, 2k overlap, 18k full-text cap
2. Scoring — each chunk → each LLM via strict JSON schema with anti-hallucination constraints
3. Span validation — post-hoc fuzzy match of every quote against the source chunk + LLM-as-judge backstop for topic relevance
4. Aggregation — confidence-weighted mean across chunks per manifesto (default; alternative aggregations supported by the website tool)

Calibration

A 50-manifesto matched-pair pilot vs Claude Sonnet 4.6 as gold-standard:

Model	mean |gap|	mean ρ
GPT-4.1-mini	0.73	0.73
Gemini Flash (latest)	0.32	0.81
Gemini 2.5 Flash + thinking	1.91	0.72

gemini-2.5-flash with explicit thinking budget was rejected (over-attributes populism). gemini-flash-latest is the chosen Google variant.

Inspection

Per-manifesto evidence browser, cross-model agreement table, and an interactive WebR-powered regression module are at sstoeckl.github.io/populism-llm/.

License

Component	License
Annotations (scores, statuses, derived metrics)	CC-BY 4.0
Verbatim quotes (≤8 words) + context (≤80 words)	Fair use for scientific commentary; cite manifesto_id
Pipeline code	MIT — github.com/sstoeckl/populism-llm
Original manifesto text	Not in this dataset — © Manifesto Project / WZB. Obtain via manifesto-project.wzb.eu

Citation

@misc{stoeckl2026populismllm,
  author    = {Stoeckl, Sebastian},
  title     = {Populism-LLM: A Cross-Validated LLM Annotation of
               European Party Manifestos},
  year      = {2026},
  publisher = {Hugging Face Datasets},
  url       = {https://huggingface.co/datasets/sstoeckl/populism-llm}
}

@article{stoeckl2026populism,
  author  = {Stoeckl, Sebastian},
  title   = {Populism and Liberalism in European Party Manifestos:
             A Cross-Validated LLM Approach},
  year    = {2026},
  note    = {Working paper; DOI: tba}
}

References

• Lehmann, P., et al. (2024). The Manifesto Project Dataset. WZB.
• Menkveld, A. J., et al. (2024). Nonstandard Errors. Journal of Finance.
• Simonsohn, U., Simmons, J. P., & Nelson, L. D. (2020). Specification curve analysis. Nature Human Behaviour.

Contact

Sebastian Stöckl · University of Liechtenstein · sebastian.stoeckl@uni.li