shevlinn/dail-pq

One Million Questions — a labelled corpus of Dáil parliamentary questions

A machine-labelled corpus of every written parliamentary question (PQ) submitted in Dáil Éireann between 2002 and 2026 — roughly 965,000 questions, each annotated for thirteen analytic dimensions including local-vs-national scope, policy-vs-administration, beneficiary, ask type, Comparative Agendas Project topic, and reply substantiveness.

The labels were produced by a self-hosted Qwen 3 32B (FP8) instance running on RunPod H100 SXM via vLLM. Validation against a 500-question comparison set labelled by Anthropic Claude Haiku 4.5, Qwen 3 27B, and Qwen 3 32B is documented in the project paper.

The underlying parliamentary record is the public Oireachtas API (https://data.oireachtas.ie/...); our contribution is the labels and the schema. Files refresh on each labelling-pipeline snapshot.

Files

File	Rows	What's in it
questions.parquet	965,039 (~36 MB)	Full corpus rows. All thirteen labels + asking-TD metadata (name, member URI, constituency, party, Dáil number, government/opposition status). No question or reply text. ZSTD-compressed Parquet.
sample-12k.jsonl	12,000 (~41 MB)	Deterministic uniform-random sample drawn from the labelled rows, with question text and minister reply. Use this for skim-reading, fine-tuning, or to validate labels by eye.
constituency-rates.csv	43 (~7 KB)	Per-(modern-)constituency rollups: total questions, labelled fraction, local share, top topic, volume by year (wide format). One row per Dáil-34 constituency.

Quickstart

Python (Hugging Face datasets)

from datasets import load_dataset

# Just the labels and TD metadata, fast.
corpus = load_dataset("shevlinn/dail-pq", "corpus", split="train")

# A 12k sample with full question text + minister reply.
sample = load_dataset("shevlinn/dail-pq", "sample", split="train")

# Per-constituency aggregates.
agg = load_dataset("shevlinn/dail-pq", "aggregates", split="train")

DuckDB (no download)

SELECT cap_topic, COUNT(*) AS n
FROM 'https://huggingface.co/datasets/shevlinn/dail-pq/resolve/main/questions.parquet'
GROUP BY 1 ORDER BY 2 DESC;

Pandas / Polars / R

Direct download links:

https://huggingface.co/datasets/shevlinn/dail-pq/resolve/main/questions.parquet
https://huggingface.co/datasets/shevlinn/dail-pq/resolve/main/sample-12k.jsonl
https://huggingface.co/datasets/shevlinn/dail-pq/resolve/main/constituency-rates.csv

Parquet is read natively by pandas (pd.read_parquet), polars (pl.read_parquet), R (arrow::read_parquet), and Stata (via the parquet package).

Schema — questions.parquet

One row per question. All label fields are drawn from closed enum sets.

Identity / metadata

Column	Type	Notes
id	string	Question URI of the form https://data.oireachtas.ie/ie/oireachtas/question/YYYY-MM-DD/pq_<n>. Primary key.
date	date	Date of the question.
year	int	Convenience column.
td_name	string	Canonical name of the asking TD.
td_member_uri	string	Oireachtas member URI. Stable across constituency / party changes.
constituency_at_date	string	The TD's constituency on date using historic boundaries.
constituency_modern	string	Same constituency mapped onto the 2024 (43-seat) boundaries. Use this for any cross-time aggregate.
party	string	The TD's party at date.
dail_no	int	Which Dáil sat (e.g. 33 = 2020–2024).
asker_govt_status	string	in_government, opposition, or support_party at date.
minister_role	string	The role the question was directed to (e.g. "Health", "Finance").
department	string	Government department of minister_role, where mapped.
labeller	string	Always qwen-3-32b-fp8 for this file.

Categorical labels

Column	Values	Approx. frequency
scope	national, local, mixed	69 / 31 / <1
policy_admin	policy, administration, mixed	32 / 68 / <1
beneficiary	general_public, single_named_place, sector_or_group, named_individual, specific_business, none	—
ask_type	request_information, request_action, propose_policy_change, score_political_point, express_grievance, request_meeting	94 / 4 / …
temporal_framing	current, historical, forward_looking	—
cap_topic	21 CAP major-topic strings (health, education, government_operations, social_welfare, agriculture, transportation, housing, law_and_crime, domestic_commerce, environment, immigration, macroeconomics, international_affairs, energy, culture, technology, defence, civil_rights, foreign_trade, labour, public_lands)	health ~20% (largest)
cap_subtopic	CAP minor topic strings; long tail; NULL ~33% of rows	—
stance	not_applicable, neutral_factual, anti_government_position, pushes_more_permissive, pro_government_position, pushes_more_restrictive	mostly not_applicable
genuine_question	genuine, rhetorical, planted_softball	99.6 / 0.3 / 0.1
response_substantiveness	substantive, referred_to_agency, will_revert, no_information_available, refused, not_applicable	78 / 20 / 1 / 1 / <1 / <1

Boolean labels

Column	Notes
monetary_claim	True if the question or expected answer makes a monetary claim. ~16% true.
urgency_signal	True if the question carries explicit urgency framing. ~1% true.

Free text

Column	Notes
brief_reason	Short labeller-written justification for its categorical choices. Mean ~120 chars. Useful as a sanity-check signal; not a label per se.

Schema — sample-12k.jsonl

One JSON object per line. Same labels as questions.parquet, plus question_text and reply_text. Nested for readability:

{
  "id": "https://data.oireachtas.ie/ie/oireachtas/question/2024-03-12/pq_19237",
  "date": "2024-03-12",
  "year": 2024,
  "td": {
    "name": "Pearse Doherty",
    "member_uri": "https://data.oireachtas.ie/ie/oireachtas/member/id/PearseDoherty",
    "constituency": "Donegal",
    "constituency_modern": "Donegal",
    "party": "Sinn Féin",
    "dail_no": 33,
    "govt_status": "opposition"
  },
  "minister_role": "Finance",
  "department": "Finance",
  "question_text": "To ask the Minister for Finance …",
  "reply_text": "I am informed by the Revenue Commissioners that …",
  "labels": {
    "scope": "national",
    "policy_admin": "policy",
    "beneficiary": "general_public",
    "ask_type": "request_information",
    "temporal_framing": "current",
    "cap_topic": "macroeconomics",
    "cap_subtopic": null,
    "stance": "neutral_factual",
    "monetary_claim": true,
    "urgency_signal": false,
    "genuine_question": "genuine",
    "response_substantiveness": "substantive",
    "brief_reason": "Asks for stats on revenue receipts; minister provides figures."
  },
  "labeller": "qwen-3-32b-fp8"
}

The sample is a deterministic uniform-random draw from the labelled corpus, seeded for reproducibility. Re-running the build with the same --seed produces the same 12k rows.

Schema — constituency-rates.csv

One row per modern (2024) Dáil constituency. Pre-2024 constituencies are mapped forward to their successor; questions asked under a historic boundary are attributed using point-in-time membership.

Column	Notes
constituency	Modern constituency name (one of 43).
total_questions	Total PQs asked by TDs from this constituency (all years).
labelled_n	How many of those have Qwen labels.
local_n, national_n, mixed_n	Counts of scope values among labelled rows.
local_share_of_labelled	local_n / labelled_n. The headline "clientelism" rate.
top_topic	Most-asked CAP major topic in this constituency.
top_topic_n	Count of the top topic.
n_2002, n_2003, …, n_2026	Total PQ volume per year.

Methodology

The labelling pipeline is open-source and reproducible. See the project repo for the prompts, the few-shot examples, the comparison-set evaluation, and the inter-labeller agreement analysis (Cohen's κ on a three-way Haiku × Qwen-27B × Qwen-32B set of 500 questions).

The corpus is the frame, not a sample: every written PQ in the public Oireachtas record between 2002-01-01 and the snapshot date is included. Some years (early-2000s) are sparser because the parliamentary calendar itself was sparser; 2026 is partial.

Caveats

1. 2026 is a partial year at the snapshot date. Trend charts should clip or annotate.
2. cap_subtopic is often null — the labeller leaves it blank when the topic is not specific enough to warrant a sub-classification (~33% of rows).
3. Rare classes (mixed, none for beneficiary, not_applicable for response_substantiveness, the non-genuine values) have <1% prevalence; pool or exclude them for statistical tests.
4. Constituency mapping: pre-2024 boundaries are mapped forward via the table in the build script. Two seats in the modern Dublin metro split ambiguously across the old Dublin South / Dublin Bay South line — those rows are mapped to the larger successor; treat sub-Dublin breakdowns as approximate.
5. Labels are model output, not ground truth. The κ on the comparison set is in the paper — use it to size your error bars.

Licence

The dataset is released under CC BY 4.0 — free to use, remix, and redistribute with attribution. The underlying parliamentary record is itself public-domain Oireachtas content; what is licensed here is the labels and the schema we layered on top.

Citation

@dataset{dail_pq_corpus,
  title  = {One Million Questions: a labelled corpus of Dáil parliamentary questions},
  author = {Shevlin, Neil},
  year   = {2026},
  url    = {https://huggingface.co/datasets/shevlinn/dail-pq},
  note   = {CC BY 4.0}
}

Contact

Corrections, methodological questions, and reuse-with-attribution notifications are welcome. Open an issue on the project GitHub repo.