OCPSG-Benchmarking-LLMs/ocpsg-silver-standard-parliamentary-speeches

OCPSG Silver Standard Parliamentary Speeches

Release: v1.0.0-rc.2 "Lively Monolith"

Dataset Summary

This repository contains the final silver-standard parliamentary speech dataset produced in the Oxford Computational Political Science Group benchmarking workflow. The dataset is intended for multilingual policy agenda classification and downstream benchmarking and fine-tuning tasks.

The release includes country-level train/validation/test splits. In the original workflow, pooled multilingual train/validation/test splits are also created, but the pooled Parquet files are not distributed in this repository in order to avoid redundancy and reduce repository size.

The pooled train/validation/test files are produced in the original workflow, and their composition is documented here for transparency, but the pooled Parquet files themselves are not uploaded to this repository. They can be reconstructed by concatenating the corresponding country-level split files for all included countries.

Current pooled dataset summary:

• Total pooled observations: 718,502
• Countries included: 23
• Unique labels (predicted_label): 19
• Train rows: 502,983
• Validation rows: 107,760
• Test rows: 107,759

Countries currently included in the pooled release:

Austria, Belgium, Bosnia and Herz., Bulgaria, Croatia, Czech Republic, Denmark, Estonia, France, Germany, Great Britain, Greece, Hungary, Iceland, Latvia, Norway, Poland, Portugal, Serbia, Slovenia, Sweden, Turkey, Ukraine

Country statuses represented in the pooled release:

include_pooled_coverage_only, include_standalone_and_pooled

Source Data

The dataset is built from two source corpora:

1. ParlaMint for most countries.
2. ParlLawSpeech for Germany.

For ParlaMint countries, the exported data include both the original-language speech field (text) and the English field (text_en) when available. For ParlLawSpeech / Germany, the dataset includes text, while text_en is left missing.

The export also includes light-normalised source metadata fields:

• date
• party
• speaker_gender

These fields are preserved as closely as possible to the original source data. Empty strings and missing values are normalised to missing. Dates are exported in YYYY-MM-DD format.

Silver-Standard Construction

The final label used in this dataset is predicted_label, obtained directly from the CWCD / entropy output. No fallback label is used. The script validates that predicted_label is complete in the final filtered silver pool before writing any splits.

Our pipeline export also retains two audit variables from early stages:

• entropy: the CWCD / entropy value associated with each observation.
• agreement_level: the maximum count of identical labels across the three baseline model outputs used to reconstruct the agreement profile.

Filtering

This release is based on the final filtered silver pool rather than the unfiltered universe.

For ParlaMint countries, observations are restricted to speeches where:

• speaker_MP == "MP"
• speaker_role == "Regular"

For Germany / ParlLawSpeech, the filtering logic follows the project’s regular-legislator approximation used in the previous stages.

Country Inclusion Logic

Country inclusion decisions are inherited from early stages of the pipeline, where countries were assessed using retained sample size, label coverage, concentration, normalised entropy, and post-filtering label survivorship.

Countries with country_status = include_standalone_and_pooled were judged strong enough to support both:

• country-level standalone train/validation/test splitting; and
• inclusion in the pooled multilingual corpus.

Countries with country_status = include_pooled_coverage_only were judged useful for the pooled multilingual corpus but not strong enough for robust standalone country evaluation.

Countries labelled review or excluded are not included in this release (i.e., Netherlands, Italy, and Spain).

Split Strategy

Country-first splitting

Splits are created within each country first. Only after the country-level train/validation/test partitions are created are the pooled multilingual splits assembled.

This means that:

• each country contributes its own train, validation, and test partitions;
• the pooled train, validation, and test files are concatenations of those country-level partitions, although those pooled Parquet files are not distributed in this repository;
• no pooled split is created before the country-level split logic is applied.

This design preserves country coherence and avoids leakage between country-specific and pooled evaluation settings.

Eligible countries

Only countries classified as:

• include_standalone_and_pooled
• include_pooled_coverage_only

are included in this release.

Countries classified as review or excluded are not part of this release in this stage.

Stratification

Within each country, observations are split by predicted_label. This means the split is label-aware inside each country and does not pool labels across countries during the splitting step.

Sparse-label safeguard

We used a tiered sparse-label safeguard in order to avoid unstable validation/test partitions for very small labels:

• If a country-label has 20 or more observations, the split uses the standard 70/15/15 allocation.
• If a country-label has 10 to 19 observations, the split assigns 1 validation and 1 test observation, with the remaining observations assigned to train.
• If a country-label has 5 to 9 observations, the split assigns 1 test observation and places the remaining observations in train.
• If a country-label has fewer than 5 observations, all observations for that label are assigned to train only.

This safeguard preserves rare labels in the released data while avoiding the creation of extremely sparse validation and test strata.

Dataset Structure

The output is organised as follows:

Root level

• README.md

Pooled train, validation, and test splits are created in the original workflow, but the corresponding pooled Parquet files are not distributed in this repository.

Country-level splits

• <CODE>/train.parquet
• <CODE>/validation.parquet
• <CODE>/test.parquet

Manifests

• manifests/split_manifest_by_country.csv
• manifests/split_manifest_by_label.csv
• manifests/split_manifest_pooled.csv

The manifest files are included as audit and reproducibility artefacts:

• split_manifest_by_country.csv summarises retained observations and label counts by country and split.
• split_manifest_by_label.csv reports the label distribution within each country and split.
• split_manifest_pooled.csv summarises the pooled train, validation, and test splits produced in the original workflow.

The pooled splits can be reconstructed downstream by concatenating the corresponding country-level train, validation, and test files across all countries included in the pooled release logic.

Data Fields

Each exported row contains the following fields:

• id: observation identifier
• country_code: two-letter country code
• country: country name
• country_status: final country-level inclusion status carried forward from the previous stage
• split: one of train, validation, or test
• predicted_label: final silver-standard label from the CWCD / entropy workflow
• entropy: CWCD / entropy value
• agreement_level: agreement count across the three baseline model outputs
• date: speech date in YYYY-MM-DD format
• party: party field preserved from source
• speaker_gender: speaker gender field preserved from source where available
• text: original-language speech text
• text_en: English speech text where available
• source_corpus: source corpus name (ParlaMint or ParlLawSpeech)

License

This dataset is released under CC BY 4.0.

Limitations

This is a silver-standard dataset rather than a fully human-coded gold standard. Labels derive from the CWCD / entropy workflow applied after the project’s silver-standard selection pipeline.

Some countries are included for pooled multilingual coverage only and are not intended for standalone evaluation. Rare labels may be present only in train because of the sparse-label safeguard described above.

The speaker_gender field is available only where present in the source corpus and may therefore be missing for some countries.

Citation

González-Bustamante, B., Bellens, T., Klamm, C., & Koch, M. (2026). OCPSG silver standard parliamentary speeches (Version v1.0.0-rc.2, "Lively Monolith") [Dataset]. Oxford Computational Political Science Group.

Contact

• Website: https://ocpsg-benchmarking-llms.github.io
• Email: ocpsg@politics.ox.ac.uk