MEDLINE-S1

MEDLINE-S1 is a set of typed decisions over PubMed abstracts. Each abstract has several questions about it: the study design, the main disease area, and yes/no questions about humans, animals, adverse effects, drug therapy and two age or sex groups. Every label comes from NLM's MeSH indexing and publication types, so no language model was involved in labelling. It is the training and evaluation set introduced in From Retrieval to Typed Decisions: Calibrated System One Models from Biomedical Sentence Encoders. Code is available at https://github.com/pritamdeka/sbert2s1.

This release contains PMIDs, questions and labels, but no abstract text. Run hydrate.py to restore the titles and abstracts from PubMed (see below).

Splits

Split	Abstracts	Decisions	Source baseline files
train	40,000	242,986	14 files of the 2026 PubMed baseline
validation	2,000	12,138	same 14 files
calibration	400	2,418	same 14 files (used only to fit temperatures)
test	5,000	30,522	3 other files, disjoint from training

manifest.json lists the exact baseline files, the MeSH trees year, the question counts and the SHA-256 of each full split file.

Questions

Question	Type	Label source	Asked for
design	choice of 6	publication types (+ observational-design MeSH); precedence systematic review/meta-analysis > RCT > other trial > case report > observational > narrative review	abstracts with an unambiguous design
disease_area	choice	the single MeSH C-tree category among the major-topic descriptors, against 3 to 7 random other categories	abstracts with exactly one such category
humans, animals	yes/no	MeSH check tags	all
adverse	yes/no	any heading with the subheading adverse effects	all
drug_therapy	yes/no	any heading with the subheading drug therapy	all
group_* (infants, children, adolescents, adults, older adults, female, male)	yes/no	MeSH check tags	2 sampled per abstract about humans

Only citations from 2005 to 2021 with manual or curated (not fully automated) MeSH indexing and an abstract of at least 400 characters are included. PMIDs in any dev or test set of the companion BioDecide suite are excluded.

Format

Each line is one abstract:

{"id": "medline_s1:test:000000", "task": "medline_s1", "split": "test", "pmid": 30365872, "year": 2019,
 "state_sha256": "a39408c1...",
 "questions": "{\"disease_area\": {\"type\": \"choice\", \"instructions\": \"Which disease area is ...\", \"criteria\": {...}}, \"humans\": {\"type\": \"noul\", ...}}",
 "gold": "{\"disease_area\": {\"probabilities\": {\"Neoplasms\": 1.0}}, \"humans\": {\"probabilities\": {\"true\": 1.0}}, ...}"}

questions and gold are JSON strings, because different abstracts have different questions. Decode them with json.loads. noul is a yes/no question with the options true and false. Gold labels are distributions over the option keys. The state, once restored, is the title, a space, and the abstract.

Restoring the abstracts

pip install lxml
python hydrate.py --data data --out prepared/medline_s1 --email you@example.org

This fetches the citations from PubMed E-utilities (about 240 requests; an NCBI API key, --api-key, makes it faster) and writes the full records in the format the sbert2s1 code reads. Every restored text is checked against state_sha256. E-utilities returns the current version of each citation, so a few titles or abstracts may have been revised since the paper's build. The script reports them, and --drop-changed leaves them out. If you have the baseline files listed in manifest.json, --baseline-dir <folder> restores an exact copy, and every split file then matches the paper's hash.

Licence and attribution

The questions, labels and split assignments are released under CC BY 4.0. The labels are derived from MEDLINE/PubMed citation data and MeSH.

This product uses publicly available data from the U.S. National Library of Medicine (NLM), National Institutes of Health, Department of Health and Human Services; NLM is not responsible for the product and does not endorse or recommend this or any other product. The data come from the 2026 MEDLINE/PubMed baseline and may not reflect the most current information available from NLM.

Abstracts are not redistributed here. Their copyright stays with the publishers or authors, and hydrate.py retrieves them from NLM under NLM's terms of use.

Citation

@misc{deka2026sbert2s1,
  title  = {From Retrieval to Typed Decisions: Calibrated System One Models from Biomedical Sentence Encoders},
  author = {Deka, Pritam},
  year   = {2026},
  note   = {arXiv preprint}
}