stojchevamarija/mk-parliament-qa

Macedonian Parliamentary QA (RAG Evaluation)

A question-answering benchmark in Macedonian, built over transcripts of plenary sessions of the Assembly of the Republic of North Macedonia. It is intended for evaluating retrieval-augmented generation systems on a low-resource language, with questions that require single-passage, multi-passage, and cross-document reasoning.

Dataset structure

Two configs are provided.

questions

Field	Type	Description
id	string	Stable identifier, prefixed by question type (SC-, MC-, MD-)
question_type	string	single_chunk, multi_chunk, or multi_doc
question	string	Question in Macedonian
answer	string	Reference answer in Macedonian
language	string	ISO 639-1 code, always mk
source_documents	list[string]	Document IDs required to answer
speakers	list[string]	Speakers whose turns contain the evidence
evidence	list[object]	Evidence spans, see below
num_evidence_spans	int	Length of evidence
num_source_documents	int	Length of source_documents
has_unresolved_speaker	bool	True if any span has an unidentified speaker

Each item of evidence:

Field	Type	Description
document_id	string	Foreign key into the documents config
speaker	string | null	Speaker name, null when not identified in the transcript
speaker_resolved	bool	False when speaker is null
speaker_verified	bool	False when the attribution was inferred rather than read off the transcript
start, end	string	HH:MM:SS offsets into the source recording
start_seconds, end_seconds, duration_seconds	int	Same offsets in seconds

documents

Field	Type	Description
document_id	string	e.g. sednica_68
session_number	int	Session number in the parliamentary term
title, title_en	string	Session title
date	string	Session date, YYYY-MM-DD
language	string	mk
source_platform	string	youtube
youtube_id	string	Video identifier
source_url	string	Link to the recording
modality	string	audio_video_transcript

Composition

Question type	Count	Definition
single_chunk	30	Answerable from one contiguous speaker turn
multi_chunk	5	Requires two or more turns within a single session
multi_doc	4	Requires evidence from both sessions
Total	39

Source sessions:

Document ID	Date	Recording
sednica_68	2025-09-12	https://www.youtube.com/watch?v=phkvPfId6j4
sednica_77	2025-12-23	https://www.youtube.com/watch?v=LvduBdsEdbk

Usage

from datasets import load_dataset

qa = load_dataset("stojchevamarija/mk-parliament-qa", "questions", split="test")
docs = load_dataset("stojchevamarija/mk-parliament-qa", "documents", split="test")

multi_doc = qa.filter(lambda x: x["question_type"] == "multi_doc")

Evidence spans can be resolved back to the recordings:

row = qa[0]
span = row["evidence"][0]
doc = {d["document_id"]: d for d in docs}[span["document_id"]]
print(f"{doc['source_url']}&t={span['start_seconds']}s")

Intended use

Evaluation of retrieval quality and answer faithfulness for Macedonian RAG pipelines. The evidence field supports retrieval metrics (recall@k, MRR) in addition to generation metrics, since the gold passages are identified by document, speaker, and time range.

Limitations and caveats

• Size. 39 questions is a diagnostic set, not a training set.
• Type imbalance. 77% of items are single-chunk, reflecting the kinds of questions the recordings support. Report per-type breakdowns.
• Abstractive answers. References are paraphrases, not extracted spans. Exact match is inappropriate; use LLM-as-judge, RAGAS-style metrics, or human evaluation.
• Provenance. Transcripts are ASR output over the linked recordings, not the official stenographic record. Timecodes refer to the recordings.
• Contested content. Answers describe what a speaker claimed during adversarial debate, including allegations against named individuals. They are not verified fact.
• Named individuals. All speakers are public officials in a public capacity and the recordings are public. Downstream use should not present claims made in debate as established statements about those individuals.

Citation

@misc{mk_parliament_qa,
  title  = {Macedonian Parliamentary QA (RAG Evaluation)},
  author = {Stojcheva, Marija and Mishkovski, Igor},
  year   = {2026},
  url    = {https://huggingface.co/datasets/stojchevamarija/mk-parliament-qa}
}