Hugging Face's logo

Datasets:
shawnzhg
/
BioLitBench

Tasks:
Text Generation
Summarization
Modalities:
Tabular
Text
Formats:
parquet
Languages:
English
Size:
1M - 10M
ArXiv:
Tags:
biomedical
literature-review
scientific-writing
benchmark
claim-graph
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
BioLitBench / README.md
shawnzhg's picture
shawnzhg
Add files using upload-large-folder tool
3ddc158 verified
preview code
|
Raw History Blame Contribute Delete
10.1 kB
metadata 
license: cc-by-4.0
pretty_name: BioLitBench
language:
  - en
task_categories:
  - text-generation
  - summarization
tags:
  - biomedical
  - literature-review
  - scientific-writing
  - benchmark
  - claim-graph
size_categories:
  - 1K<n<10K
configs:
  - config_name: tasks
    default: true
    data_files:
      - split: train
        path: viewer/tasks/train.parquet
      - split: test
        path: viewer/tasks/test.parquet
  - config_name: claims
    data_files:
      - split: train
        path: viewer/claims/train.parquet
      - split: test
        path: viewer/claims/test.parquet
  - config_name: relations
    data_files:
      - split: train
        path: viewer/relations/train.parquet
      - split: test
        path: viewer/relations/test.parquet
  - config_name: references
    data_files:
      - split: train
        path: viewer/references/train.parquet
      - split: test
        path: viewer/references/test.parquet

BioLitBench

Try SCRIBE live at glkb.org

SCRIBE, the literature-review agent of this paper, runs live at glkb.org. Try it on your own research question.

Paper Code Try it

BioLitBench is the benchmark of the paper What Can a Leaderboard Certify? Compositional Controllability for Fair Evaluation and Training of Biomedical Literature-Review Agents (Zhaowei Han, Xiang Zhang, Lingxiao Guan, Danqi Hu, Kai Liu, Kevin Chang, Jie Liu; University of Michigan).

Each task asks a system to write a literature review on the topic of a published biomedical review article, using only literature published up to the task's cutoff year. The human review is the reference: it is represented as a claim graph (its claims, the references each claim cites, the typed relations between claims, hard negatives and the section hierarchy), and generated reviews are scored claim by claim against it.

Try SCRIBE on glkb.org

SCRIBE, the literature-review agent introduced in the paper, is deployed at glkb.org. Ask a biomedical research question and it writes a literature review with citations to the biomedical literature. No installation is needed.

Leaderboard

Capability scores under fixed input (every system receives the review's reference list) at the primary observation point, as reported in the paper. Scores are fractions of human level; Rank is the certified rank interval. Within each column, the highest value is bold and the second highest is underlined.

Backbone System Composite Rank Synthesis Reasoning Planning Writing Form
Qwen3.8-27B SCRIBE (ours, try it) 0.606 1–2 0.689 0.594 0.319 0.613 0.814
Qwen3.8-27B SCRIBE (untrained) (ours) 0.583 1–3 0.686 0.601 0.276 0.565 0.786
gpt-5.6-luna SCRIBE-Luna (ours) 0.564 2–6 0.700 0.611 0.177 0.568 0.767
gpt-5.6-luna SurveyG 0.516 4–12 0.568 0.541 0.272 0.509 0.689
gpt-5.6-luna OpenAI tool loop 0.515 4–12 0.741 0.552 0.136 0.487 0.660
gpt-5.6-luna SurveyGen-I 0.496 6–13 0.591 0.510 0.182 0.449 0.746
gpt-5.6-luna AutoSurvey 0.495 6–13 0.638 0.533 0.135 0.483 0.687
gpt-5.6-luna LiRA 0.484 4–13 0.539 0.495 0.254 0.442 0.716
gpt-5.6-luna LLM×MR 0.469 10–13 0.537 0.515 0.182 0.446 0.666
claude-sonnet-5 Claude Science 0.541 3–10 0.700 0.603 0.160 0.507 0.736
claude-sonnet-5 Claude Code 0.534 3–12 0.650 0.547 0.185 0.587 0.703
proprietary Elicit SR§ 0.535 3–10 0.668 0.526 0.185 0.570 0.723

Published pipelines use their reference-fed configurations. §Elicit Systematic Review; it reads full text and cites at most 80 papers. Ranks are computed over all systems evaluated, including one commercial system that is not reported. n = 50 tasks in 23 subfield clusters. The same-pool leaderboard (every system searches the same frozen corpus) and the full-length results are in the paper's appendix.

Dataset

train test
tasks (review articles) 1,842 50
claims 220,968 5,740
review-level relations 748,047 18,691
hard negatives 91,541 2,856
cited references 306,158 7,344
  • test is exactly the 50 tasks of the paper's leaderboard. train holds the 1,842 training articles. There is no validation split.
  • This dataset contains no article text and no abstract text. It ships the claim graphs and annotations with every text value replaced by a sha256 digest or a locator, plus identifiers, bibliographic metadata and the task definitions. The text is rebuilt locally from PMC and PubMed with the scripts in the code repository, and every rebuilt file is checked against its digest.

Included

  • Tasks: one TaskSpec per task (topic, cutoff year, target length), the gold record (review PMID, cited PMIDs) and, for the fixed-input track, the evidence bundle and reference allowlist of each task.
  • Claim graphs and annotations: the claim graph and section structure of every review (dehydrated/, text replaced by digests and locators), reference lists with PMID, DOI and title, and grounding scores.
  • Splits (splits/): train and test id lists, per-task metadata, the 399 test peers and the training articles excluded from evaluation.
  • Manifests (manifests/): the digest of every file to rebuild, abstract digests, pinned PMC article versions, release transforms, per-article licences and citations.
  • Viewer tables (viewer/): text-free tables for the dataset viewer above (tasks, claims, relations, references).

Not included

  • No review-article text and no abstract text.
  • No system outputs and no score tables. Scoring is done with the code repository; the frozen scoring parameters of the test split are part of the code (evaluation/frozen/test).
  • No retrieval pool: v1.0 supports the fixed-input track only.
  • The 150 held-out articles of the paper that are not test tasks are not released. They were not used for training or model selection.

Usage

Download the dataset, rebuild the text and prepare a data root with the code repository, then run generation, scoring or training as described in its README (section Data):

huggingface-cli download shawnzhg/BioLitBench --repo-type dataset --revision v1.0 --local-dir <hf>
git clone https://github.com/shawnzhg/LitReview-SCRIBE && cd LitReview-SCRIBE
python biolitbench/release/fetch_articles.py   --snapshot <hf> --profile test --work <work> --email <email>
python biolitbench/release/parse_jats.py       --xml <work>/xml --out <work>/articles
python biolitbench/release/fetch_abstracts.py  --snapshot <hf> --profile test --work <work> --email <email>
python biolitbench/release/rehydrate.py        --snapshot <hf> --root <root> --profile test --articles <work>/articles --abstracts <work>/abstracts.jsonl
python biolitbench/release/prepare_root.py     --snapshot <hf> --profile test --root <root>
python biolitbench/release/verify_data.py      --root <root> --profile test --snapshot <hf>
  • --profile test builds the test root (the 50 test tasks and their 399 peers); --profile train builds the training root.
  • The fetch scripts follow the NCBI usage rules: at most 3 requests per second (10 with --api-key), and long request series only on weekends or 21:00–05:00 US Eastern time.
  • PubMed records change over time. If an abstract no longer matches its digest, rehydrate writes today's text and records it in rehydrate_report.json; results on such text are marked and are not official scores.
  • Training on the test peers. splits/train_eval_excluded.txt lists 403 training articles (the 399 test peers and 4 reviews cited by test reviews). Training on them makes test scores optimistic; the SCRIBE recipe excludes them.
  • The rebuilt text stays under the copyright of its publishers and authors and the PMC and NLM terms. Keep it local and do not redistribute it.

Licences

  • Our annotations (identifiers, labels, locators, digests and splits) are licensed under CC BY 4.0.
  • Bibliographic metadata (titles, authors, journals, years, PMIDs, DOIs) comes from NLM databases (PubMed, PMC) under the NLM Terms and Conditions and is not relicensed.
  • OpenAlex labels (topic, subfield, field) come from OpenAlex, which releases its data under CC0.
  • Review articles: manifests/licences.csv gives each review's licence; manifests/citations.csv gives authors, title, journal, year, DOI, PMID and copyright holder and is the attribution record for every review whose annotations are included.

Courtesy of the U.S. National Library of Medicine. NLM is not responsible for this dataset and does not endorse it.

Citation 

@misc{han2026leaderboardcertifycompositionalcontrollability,
      title={What Can a Leaderboard Certify? Compositional Controllability for Fair Evaluation and Training of Biomedical Literature-Review Agents},
      author={Zhaowei Han and Xiang Zhang and Lingxiao Guan and Danqi Hu and Kai Liu and Kevin Chang and Jie Liu},
      year={2026},
      eprint={2609.32318},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2609.32318},
}