ResPolitica

A unified dataset for political information, including from Voting Advice Applications (VAAs)

---

Information Mediation - Tasks & Types

ResPolitica currently covers party-position mediation ($T_{\text{party-pos}}$): given a party and a policy statement, retrieve the party's stance and rationale. This is one task type within the broader political information mediation family:

Task Type	Example
Party-Position (current scope)	"What is [Party]'s position on rent control?"
Navigation	"Which parties support rent control?"
Comparison	"How do Party A and Party B differ on migration policy?"
Issue Mapping	"What are the main political conflicts around housing policy?"

Future versions will expand beyond VAA data to cover additional political information types (e.g. candidate information, debates, manifestos candidate interviews, parliamentary votes) and mediation tasks as grounding data and evaluation criteria are developed.

---

Data Sources

Source	Coverage
🇩🇪 Wahl-O-Mat (qual-o-mat-data)	DE federal + state elections, 2017–2025
🇳🇱 StemWijzer	NL TK2025 only
🇪🇺 EUANDI 2024	EU Parliament 2024 (DE, ES, FR, GR, IT + EP groups)

Note: StemWijzer data for TK2017, TK2021, and TK2023 is excluded for copyright reasons.

---

Dataset Files

All files under datasets/processed/.

File	Description	Rows
respolitica_unified.parquet / .csv	All observations, all sources, all fields	12,584
respolitica_unified_federal.parquet / .csv	Federal elections only	4,444
respolitica_unified_states.parquet / .csv	State elections only	7,030
respolitica_normalized/elections.parquet	Election registry	19
respolitica_normalized/parties_canonical.parquet	Global party registry	143
respolitica_normalized/parties_election.parquet	Party × election appearances	344
respolitica_normalized/statements.parquet	Unique statements per election	666
respolitica_normalized/observations.parquet	Fact table (FKs + stance + rationale)	12,584

The respolitica_unified* files are flat denormalized exports. The respolitica_normalized/ tables are the relational decomposition of the same corpus.

---

Schema — respolitica_unified

Column	Type	Description
observation_id	str	Unique row identifier
source_dataset	str	qual-o-mat, euandi, or stemwijzer
election_id	str	Slug, e.g. bundestagswahl2025, nl_tk2025
election_date	date	ISO date
election_level	str	federal, state, or european
country_iso	str	ISO 3166-1 alpha-2
party_id	str	Canonical party slug
party_name_short	str	Short name
statement_id	str	Unique statement identifier
statement_text	str	Statement in source language
statement_text_en	str	English translation
stance_label	str	Agree, Neutral, or Disagree
stance_numeric	float64	1.0 / 0.0 / -1.0
rationale_text	str	Party rationale in source language
rationale_text_en	str	English translation
quality_flag	str	e.g. missing_rationale, clean
has_rationale	bool	False when rationale is blank or placeholder

---

Installation

git clone <repo-url>
cd ResPolitica
uv venv && uv sync

# Dataset summary
python scripts/respolitica_data.py info

# Validate data quality
python scripts/respolitica_data.py validate

# Query by party or election
python scripts/respolitica_data.py query --election bundestagswahl2025 --party CDU
python scripts/respolitica_data.py query --country nl --format json

---

Generating the Full Dataset

The pipeline parses raw VAA source files and produces the processed outputs above. It also generates the Information Environments (IEs) used in the Polistemics benchmark.

# Fetch upstream source data (Qual-o-Mat, EUANDI)
python scripts/respolitica_data.py fetch --source qualomat
python scripts/respolitica_data.py fetch --source euandi

# Parse and unify all sources
python scripts/respolitica_data.py prepare

Benchmark construction (phases 1–3) generates respolitica_augmentations.jsonl and respolitica_benchmark.jsonl:

# First generate the scope CSV (required by assemble):
python scripts/pipeline.py --stages scope
python scripts/benchmark_cli.py standardize
python scripts/benchmark_cli.py standardize-phase2
python scripts/benchmark_cli.py assemble
python scripts/benchmark_cli.py validate-output

Set benchmark.standardization.model in config.yaml to a valid LiteLLM model string before running.

---

Credits & Licenses

Wahl-O-Mat / qual-o-mat-data

• Source: Bundeszentrale für politische Bildung (BpB), Germany
• Repository: gockelhahn/qual-o-mat-data
• Original tool: wahl-o-mat.de

StemWijzer

• Source: ProDemos — House for Democracy and the Rule of Law, Netherlands
• Original tool: stemwijzer.nl

EUANDI 2024

• Source: European University Institute (EUI), Robert Schuman Centre
• Dataset: coastalcph/euandi_2024
• License: CC BY-NC-SA 4.0 — academic use with attribution, non-commercial only

---

Citation

@misc{polistemics2026,
  title   = {Polistemics: Evaluating LLMs as Information Mediators in Politics \& Elections},
  author  = {Peters, Baran},
  year    = {2026},
  note    = {Preprint}
}

License

See LICENSE.