JJBauer/bioasq-14b

BioASQ 14B — Biomedical Question Answering

The complete training corpus for BioASQ Task 14B, repackaged as a HuggingFace dataset. 5,729 expert-annotated biomedical questions, each with relevant PubMed documents, extracted snippets, an exact answer and several reference long-form answers.

This is a faithful repackaging of training14b.json — no questions were added, removed or reworded. Two columns were derived to make the challenge history usable (asq_challenge and split_13b), and exact_answer was normalized into a single consistent shape.

Splits

The dataset is split by question type rather than by train/test, because all 5,729 questions are 14B training data. Choose your own evaluation split - see below.

Split	Questions
factoid	1,695
list	1,130
summary	1,363
yesno	1,541
total	5,729

Columns

Column	Type	Description
id	string	Official BioASQ question id, stable across challenge editions
type	string	factoid, list, summary or yesno (same as the split name)
question	string	The question text
answer	list[list[string]]	Exact answer. Outer list holds distinct answer items, inner list holds accepted synonyms. Empty for summary.
ideal_answer	list[string]	Reference long-form answers, several per question
documents	list[string]	PubMed URLs of the relevant articles
snippets	list[struct]	Relevant text spans, with text, document, beginSection, endSection, offsetInBeginSection, offsetInEndSection
asq_challenge	int32	Challenge edition (1–13) in which the question was first introduced
split_13b	string	train or test — which side of the BioASQ 13 challenge split this question fell on

The answer column

BioASQ stores exact_answer in three different shapes depending on question type. This dataset normalizes all of them to list[list[string]] so the column has one Arrow type:

Type	Original	Normalized
yesno	"yes"	[["yes"]]
factoid	["Bazex syndrome"]	[["Bazex syndrome"]]
list	[["epidermal growth factor"], ["betacellulin"]]	unchanged
summary	absent	[]

For yesno, read row["answer"][0][0]. For factoid, row["answer"][0] is the list of accepted synonyms. For list, each element of row["answer"] is one required list item with its synonyms.

The asq_challenge column

The edition in which a question first appeared, computed by scanning every training release (2b–14b) and every golden test set (Task2B–Task13B), plus the BioASQ 1 task1b phase testsets that ship inside the BioASQ-training2b folder. The 301 questions marked edition 1 are the original BioASQ 1 questions.

Edition	1	2	3	4	5	6	7	8	9	10	11	12	13
Questions	301	460	494	492	498	499	499	499	492	485	330	340	340

Useful for time-based splits: training on asq_challenge <= 11 and evaluating on asq_challenge >= 12 reproduces a realistic temporal generalization setup. For the latest available train/test split (as of 16.09.2026 this is version 13b) you can also use feature split_13b.

The 13b_split column

BioASQ folds each year's golden test questions into the following year's training release, so training14b.json is exactly training13b.json (5,389 questions) plus the 340 questions used as the BioASQ 13 test set. This column records that boundary, letting you reproduce the official 13B split without downloading the golden test files.

Usage

from datasets import load_dataset, concatenate_datasets

ds = load_dataset("JJBauer/bioasq-14b")

# One question type
factoid = ds["factoid"]

# Everything at once
everything = concatenate_datasets([ds[t] for t in ["factoid", "list", "summary", "yesno"]])

Reproduce the official BioASQ 13 split:

train = everything.filter(lambda r: r["13b_split"] == "train")   # 5,389
test = everything.filter(lambda r: r["13b_split"] == "test")     # 340

A temporal split by challenge edition:

train = everything.filter(lambda r: r["asq_challenge"] <= 11)
test = everything.filter(lambda r: r["asq_challenge"] >= 12)

A note on evaluation

Because BioASQ absorbs old test sets into new training releases, the published golden test sets for editions 2–13 are almost entirely contained in this corpus: of 5,486 pooled test questions, 5,483 (99.9%) are present here. Do not evaluate on downloaded TaskNBGoldenEnriched files while training on this dataset — use 13b_split or asq_challenge to carve out a holdout instead, or evaluate against the BioASQ 14 test sets, which are not part of this corpus by construction.

Citation

If you use this dataset, you must cite the original BioASQ papers (the first two below).

@article{krithara2023bioasq,
  title   = {BioASQ-QA: A manually curated corpus for Biomedical Question Answering},
  author  = {Krithara, Anastasia and Nentidis, Anastasios and Bougiatiotis, Konstantinos and Paliouras, Georgios},
  journal = {Scientific Data},
  volume  = {10},
  pages   = {170},
  year    = {2023},
  publisher = {Nature Publishing Group UK London}
}

@article{tsatsaronis2015overview,
  title={An overview of the BIOASQ large-scale biomedical semantic indexing and question answering competition},
  author={Tsatsaronis, George and Balikas, Georgios and Malakasiotis, Prodromos and Partalas, Ioannis and Zschunke, Matthias and Alvers, Michael R and Weissenborn, Dirk and Krithara, Anastasia and Petridis, Sergios and Polychronopoulos, Dimitris and others},
  journal={BMC bioinformatics},
  volume={16},
  number={1},
  pages={138},
  year={2015},
  publisher={Springer}
}

Licensing

The BioASQ data is distributed by the BioASQ organisers; consult bioasq.org and the participants area for the terms that apply to the original corpus. This repackaging does not change those terms.