MehdiAstaraki/un-generated-qa

Multilingual generated QAs and source corpora

Multilingual supporting evidence

The qas configuration now includes native nested evidence_all_languages and evidence_alignment_status fields with the five keys en, de, es, fr, and zh. Each language contains a list of span/status objects; an unavailable language has an empty list. Original questions, answers, metadata, and corpus records are preserved. Arabic and Russian corpus versions remain available where previously published; this alignment run did not create evidence in those languages.

For easy reading, evidence_en, evidence_de, evidence_es, evidence_fr, and evidence_zh contain lists of the exact passage strings. A corpus-specific readable CSV contains these fields alongside questions and answers. In its readable language columns, one span is plain text, no span is an empty cell, and multiple spans use a JSON list. Native Parquet columns always remain lists. No span strings are joined, translated, or normalized.

Original supporting text is marked source_exact. Generated spans are literal matches in the linked target document and remain pending_review: exact quotation does not establish semantic equivalence or independent human review. The language-specific status lists retain unsuccessful/uncertain alignments and their provenance. evidence_alignment_eligibility explains excluded QAs (for example, missing exact original support or a quarantined language label); such questions remain in the dataset without invented evidence. The contract is cross_language_evidence_terra_v1. Legacy evidence/evidence_language columns and the separate evidence configuration keep their previous interpretation.

from datasets import load_dataset

qas = load_dataset("MehdiAstaraki/un-generated-qa", "qas", split="train")
row = next(row for row in qas if row["evidence_all_languages"]["en"])
print(row["question"])
print(row["evidence_all_languages"]["en"][0]["text"])
print(row["evidence_en"])

This update adds supporting passages only. It creates no retrieval chunks or new relevance judgments, and does not change source_qrels.

Full local corpus for the configured languages; this is not a claim of all upstream data.

Corpus languages: ar, en, es, fr, ru, zh.

UN document selection: resolutions and decisions; meeting records; letters and notes verbales. Uses the generation pipeline's document-type classifier and excludes agenda stubs, journals and corrigenda. Includes whole documents regardless of generation passage quality, context-length or reference-completeness limits. This is the complete selected document-type collection, not every UN document in the upstream six-way corpus.

All supplied generator candidates are retained, including low-scoring and unmatched evidence. The train split is a storage convention; no leakage-safe evaluation split has been assigned. Before splitting, group questions from the same article/act, UN document, or patent and its translation partners to avoid source and generator leakage.

Configurations

• qas: 4,691 rows

• evidence: 4,691 rows

• source_qrels: 28,146 rows

• corpus: 88,134 rows

• qas: questions, answers, raw evidence, persona/mode, model, scores, source IDs and involvement. document_text embeds the complete primary source unit; additional involved units are linked through involved_source_unit_ids to corpus. metadata_json preserves the complete input row, including persona-specific fields, prompts, grading responses, routing decisions and generation references. provenance_json identifies the input file and row/task. CSV metadata preserves its original string types.

• evidence: canonical supporting text: explicit evidence where present, otherwise answer. origin distinguishes the two. start and end refer to the original corpus unit's text, in Unicode code points, end-exclusive; only the first exact match receives offsets. Matching checks only presence (exact or whitespace-normalized), not factual entailment. Support is checked against the primary source, not every involved/translated document.

• corpus: _id, text, language, source/parent IDs, unit type and original metadata. EUR-Lex units are full articles (or explicitly involved annexes); patent units are title + abstract; UN units are entire documents joined from the original aligned lines. UN block IDs/ranges survive in QA metadata only. EUR-Lex duplicate ELI/language records use the last occurrence, matching the generation loader. Source text and patent markup are retained.

• source_qrels: positive source-unit associations derived from generation provenance, plus existing language versions of each legal unit / directly accepted chemistry translation partners. They are not manually judged relevance or chunk-level qrels. score=1 indicates an association; absent judgments do not establish irrelevance. Supplied contextual references alone do not create positives. Chemistry partners are not expanded transitively.

involved_source_unit_ids includes the primary source and explicitly involved EUR-Lex units; involved_article_ids includes article ELIs only. multi_article is true when multiple articles are resolved or the generator declared it. multi_article_declared retains that declaration; multi_article_conflict identifies a disagreement with the resolved article count. cross_act means involvement across acts; multi_source_unit also covers an article plus an annex. A translation into another language does not count as another involved article. UN and chemistry have no article-involvement annotation, so their multi_article is false and multi_article_declared is null; supplied UN references are retained only as metadata.

These source units have not yet been chunked. Create retrieval chunks with a pinned tokenizer (e.g. BAAI/bge-m3), preserve parent IDs and character ranges, and align evidence to those chunks. Do not mark every chunk of a source document relevant. Unmatched or translated evidence needs separate alignment/review before it can establish chunk-level judgments.

manifest.json records export inputs and SHA-256 hashes; report.json records coverage, models, personas, evidence matching and multi-article statistics. No upstream license is assigned to generated material by this exporter.

Attribution and source terms

Source: United Nations. Ziemski, Junczys-Dowmunt, and Pouliquen (2016), The United Nations Parallel Corpus, LREC. Corpus terms: https://conferences.unite.un.org/UNCorpus.