Note ELI5 — long-form QA from Reddit. 323,136 training queries.

Note AG News — news title to description. 715,929 training queries.

Note NPR — news title to body. 568,581 training queries.

Note CC-News — news title to article. 489,515 training queries.

Note SimpleWiki — Wikipedia to Simple Wikipedia. 101,784 training queries.

Note SearchQA — Jeopardy questions over web snippets. 117,220 training queries.

Note Quora duplicates — symmetric, with the source's own non-duplicates as negatives. 86,169 training queries.

Note AllNLI — symmetric NLI pairs, with the source's own contradictions as negatives. 282,442 training queries.

Note MS MARCO — the training source is this repo's train split (502,939 queries); the corpus is the test split, shared by every split.

Note HotpotQA — the training source is this repo's train split (85,000 queries); the corpus is the test split, shared by every split.

Note FEVER — the training source is this repo's train split (109,810 queries); the corpus is the test split, shared by every split.

Note Index of the 15 training sources across these 13 repositories.

Note THREE sources in one repo: Natural Questions, TriviaQA and SQuAD as DPR built them, as the config prefixes nq-, trivia- and squad- over one shared 21M-passage Wikipedia corpus.

Note PAQ — synthetic Wikipedia QA. 64.4M training queries over 9.0M passages; cap it before use.

Note Magicoder OSS-Instruct — code: coding problem → solution. 59,994 training queries.

Note CoRNStack Python — code: query → Python function. 59,983 training queries.

Note Gretel Synthetic Text-to-SQL — code: question → schema and SQL. 30,000 training queries.

Note PubMedQA — medical: biomedical question → abstract. 49,765 training queries.

Note MIRIAD — medical: medical question → paper passage. 60,000 training queries.

Note MedMCQA — medical: medical exam question → explanation. 29,998 training queries.

Note SciRepEval Search — science: search query → clicked paper. 48,965 training queries.

Note S2ORC title–abstract — science: paper title → abstract. 39,996 training queries.

Note Investopedia — finance: finance question → answer passage. 60,000 training queries.

Note TAT-QA — finance: question → report table and text. 12,666 training queries.

Note Financial QA 10-K — finance: question → 10-K passage. 5,563 training queries.

Note CLERC — legal: case paragraph → cited case. 50,000 training queries.

Note CaseHOLD — legal: citing context → holding. 39,996 training queries.

Note Law Stack Exchange — legal: legal question → answer. 24,307 training queries.

Note GerLayQA — legal (German): German lay legal question → lawyer's answer. 23,923 training queries.

Note RLHN SciDocs-RR — science: paper title → related paper. 12,594 training queries.

Note RLHN FiQA-2018 — finance: financial question → answer post. 5,159 training queries.

Note RLHN ArguAna — argument: argument → counter-argument. 2,508 training queries.