ubaada/en-wiki-synth-qa-50k

Wiki English QA Synthetic (Vital Articles Level 5)

Synthetic, article-grounded question–answer pairs generated from English Wikipedia. Articles are drawn from the community-curated Wikipedia:Vital articles / Level 5 list (the broadest importance tier, ~50k topics), so coverage is deliberately high-importance rather than random.

Stats

Metric	Value
Articles (rows)	49,268
QA pairs	927,490
QA pairs / article	6–64 (mean ≈ 18.3)
Generators	qwen3.6-35b-iq4xs-wikiqa (32,836 rows), deepseek/deepseek-v4-flash (16,432 rows)

Constant for every row: schema_version=wiki_qa_v1, prompt_version=prompt_final, source.dataset=wiki, source.selection_mode=vital_level_5.

Schema

One row per article.

column	type	description
record_id	string	stable unique id for the article record
model	string	LLM that generated the record's QAs
title	string	Wikipedia article title
url	string	canonical article URL
topic	string	Vital-articles Level-5 sublist the article came from (topic label, e.g. History)
qa_pairs	list<struct>	generated Q&As, see below

qa_pairs item:

field	type	description
question	string	generated question
answer	string	reference answer
difficulty	string	easy / medium / hard
question_type	string	mechanism / cause_effect / comparison / chronology / significance / detail_synthesis / definition
support	string	article_grounded / article_plus_general_context
evidence	string	source snippet backing the answer

Generation

• QA count scales with excerpt length; __QA_RANGE_INSTRUCTION__ is rendered per article from the min/target pair count below.

  excerpt length	min pairs	target pairs
  < 2,000 chars	6	9
  2,000–3,999	8	12
  4,000–5,999	10	15
  6,000–7,999	12	18
  ≥ 8,000	16	24

  (target ≈ 1.5× min.)

• Answers are constrained to the cleaned excerpt.

• Post-validation: rows below the minimum are rejected; answers shorter than 45 chars and questions shorter than 18 chars are dropped; questions containing "article/excerpt/passage/text" are dropped; each article is capped at 64 pairs.

Full prompt template (prompt_final)

Exact system prompt sent to the generator model. __QA_RANGE_INSTRUCTION__ is rendered per article from the min/target pair count above, e.g.:

Generate at least 16 interesting question-answer pairs from the article excerpt below; 24 is a suggested target. The lower bound of question/answer count must be followed, however the upper bound is a suggestion and you can generate as many questions as you like if, in your judgement, it requires more questions to cover all the fact and knowledge of the article.

The article itself is passed in a separate user message:

Title: <article title>
URL: <article url>

Article excerpt:
<<<ARTICLE>>>
<cleaned excerpt>
<<<END_ARTICLE>>>

Template text (prompt_final.txt):

You are building a high-quality instruction-tuning dataset from Wikipedia.

__QA_RANGE_INSTRUCTION__ Optimize for preserving as much of the article's knowledge as possible when converting prose into QA form. The QA set should cover most major facts, relationships, chronology, definitions, claims, distinctions, and notable details in the excerpt, while still producing questions a curious reader would genuinely ask.

Question guidelines:
- Make every question self-contained; include the relevant entity name instead of saying "this".
- Mix fact-recall, explanatory, conceptual, comparative, chronological, and thought-provoking questions.
- Prefer synthesis questions about why something mattered, how parts relate, what changed over time, what distinguishes two things, or what a detail reveals.
- Include at least 3 medium or hard questions that require connecting multiple facts from the excerpt.
- Include smaller factual questions when needed to preserve names, places, roles, definitions, classifications, dates, works, events, or measurements that would otherwise be lost.
- Related outside context is allowed only when it is clearly helpful, likely correct, and tightly connected to the article topic; label it with support="article_plus_general_context". Do not invent unsupported specifics.
- Avoid yes/no questions, fill-in-the-blank questions, citation/markup questions, and questions whose answer is only a date or a single name.

Answer guidelines:
- For article-grounded answers, use only information supported by the excerpt.
- For broader-context answers, start from the excerpt and add only cautious, relevant general knowledge.
- Answer in 2-5 sentences with concrete details.
- If the excerpt gives uncertainty or competing claims, preserve that uncertainty.
- Do not mention "the article", "the text", "the passage", or "the excerpt".

Return strict JSON only, no markdown:
{
  "qa_pairs": [
    {
      "question": "...",
      "answer": "...",
      "difficulty": "easy|medium|hard",
      "question_type": "mechanism|cause_effect|comparison|chronology|significance|detail_synthesis|definition",
      "support": "article_grounded|article_plus_general_context",
      "evidence": "short exact phrase from the excerpt"
    }
  ]
}

Usage

from datasets import load_dataset

ds = load_dataset("<user>/wiki-qa-vital5", split="train")
print(ds[0])