license: cc-by-4.0
task_categories:
- question-answering
- text-retrieval
language:
- en
tags:
- scientific-literature
- citation-prediction
- biochemistry
- chemistry
- biotechnology
size_categories:
- n<1K
BioChemCite
120 citation-prediction questions in biochemistry, chemistry and biotechnology. Given a sentence from a paper with its citation blanked out, name the paper being cited.
Built in the shape of CiteME, for the same reason and in a different literature.
Why this exists
CiteME measures finding rather than reading, which is the harder and less measured half of a research agent. But its excerpts are machine-learning papers, the densest region of any frontier model's training. A model can recognise the citation and search only to confirm it, which measures recall dressed as retrieval.
These questions sit in wet-lab and process literature, where that advantage is much weaker and where the answer has to be found rather than remembered.
Where the answers come from
Nowhere expensive. JATS articles link every in-text citation to its reference, and the reference carries a DOI:
<xref rid="B30" ref-type="bibr">30</xref> → <ref id="B30"> … 10.1016/j.…
So the ground truth ships with the question. No annotation, no LLM judge, no human labelling — and therefore nothing to disagree with.
Fields
id |
dq0000 … |
excerpt |
the sentence, with the citation replaced by [CITATION] |
target_doi |
DOI of the cited paper — the answer |
target_title |
title of the cited paper |
source_pmcid |
the article the excerpt came from |
source_license |
always cc by |
How each question was filtered
Roughly one sentence in twenty survives:
- Exactly one citation. "as shown previously [4,7,12]" has three right answers and no way to say which was meant.
- No author named. "Li et al. [CITATION]" hands over the answer. CiteME obfuscates these; here they are dropped.
- Makes a claim. Methods sections cite reagent suppliers in passing. An early build produced "polymerase was obtained from Vazyme [CITATION]" pointing at a Gibson assembly paper — unfindable, and worthless as a question. Methods, Materials and Protocol sections are excluded outright.
- 120–600 characters. Long enough to carry a specific claim, short enough to stay a citation task rather than a reading one.
- Resolves to a DOI, so scoring is exact rather than a title-matching argument.
Capped at four questions per source article, so no single paper dominates.
Licensing
Every source article is CC BY, verified individually against Europe PMC rather than assumed. This matters: about half the open-access literature in these fields is CC BY-NC-ND, and lifting a sentence out of an article and blanking its citation makes a derivative, which ND forbids. Filtering happens at source selection, not afterwards.
source_pmcid is on every row so each excerpt traces back to the article it
came from.
Known limitations
- Recency. Source articles are recent Europe PMC deposits, so the cited papers skew toward work published before them. Nothing here tests finding very new work.
- Europe PMC only, which means life-sciences-indexed venues. Chemistry published outside PMC is under-represented, and conference proceedings are absent entirely.
- Difficulty is not calibrated. Some excerpts describe the cited work precisely; others gesture at it. There is no per-question difficulty label and no human baseline — unlike CiteME, which has one at 69.7%.
- One citation per question by construction, so nothing here tests disentangling a multi-citation claim.
Building it yourself
The generator is at
benchmarks/build_domainqa.py in
the Vela repository. Re-running it produces a different draw from current
Europe PMC content rather than this exact set.