A leakage-controlled benchmark for evaluating LLMs on systematic-review / meta-analysis relevance prioritization: given a target evidence synthesis and a candidate paper, decide whether the paper should remain in a high-recall pool for subsequent practitioner screening.

Built and maintained with meta-screening-benchmark, the companion Python package used to construct and audit this dataset.

Dataset summary

Each row is a (target meta-analysis, candidate paper) pair. The candidate paper's full text (title, abstract, and PMC Open Access section text) is paired with a screening question derived from the target meta-analysis, and labeled:

label = 1: the candidate paper was cited by the target review (relevance proxy)
label = 0: the paper was not cited and is a temporally eligible topical hard negative

Every successfully resolved citation remains positive, whether or not the paper was ultimately pooled. Label-0 rows are not random negatives: they are topically related, non-cited papers retrieved from a publication-bounded PubMed search. These labels support high-recall relevance prioritization, not autonomous or expert-adjudicated final study eligibility.

The dataset spans three topics, only two of which are cardiology-specific:

Topic	Scope
`preventive_cardiology_risk`	Cardiovascular risk prediction and coronary calcium
`cardio_oncology`	Cardiotoxicity from cancer therapy
`pharmacology`	General drug safety / adverse drug reactions / pharmacovigilance (not cardiology-restricted)

Target meta-analyses have no recorded electronic or print publication date before 2024-01-01. This rule reduces the risk that the target review itself appeared in a model's pretraining corpus, but it is not a guarantee for models with later or undisclosed training cutoffs.

Supported tasks

Binary classification: predict label from the candidate text and the target screening question.
Ranking/retrieval: rank candidates within a meta_id group by predicted relevance.

Dataset structure

Data splits

Split	Rows	Label 0	Label 1	Target reviews (`meta_id`)
`train`	24,379	22,038	2,341	143
`validation`	4,790	4,358	432	30
`test`	4,270	3,866	404	30

33,439 rows total across 203 target evidence syntheses, comprising 30,262 temporally eligible hard negatives and 3,177 cited relevance proxies. Splits are grouped by meta_id (never split within a target review) and candidate-PMID overlap is pruned across splits, so the same candidate paper never appears in more than one split. See docs/DATASET.md for the full split-construction policy.

Release files

File pattern	Role	Use for models?	Open in Excel?
`*_clean.parquet`	Canonical 96-column split with complete text	Yes	No; use Python or Hugging Face
`*_clean_excel_preview.csv`	Public same-row/same-column QA copy; only declared long-text cells may be shortened	No	Yes
`manual_debug_100_rows.csv`	Deterministic 100-row sample from the previews	No	Yes

The three Parquet files are the publication snapshot and model inputs. The three preview CSVs and manual sample are also public Hugging Face artifacts, but are review aids only. Legacy *_clean_flat.csv files are not part of this release.

Data instances

{
  "meta_id": "3",
  "topic_id": "preventive_cardiology_risk",
  "target_meta_title": "Cardiovascular disease risk communication and prevention: a meta-analysis.",
  "research_question": "What is the effectiveness of cardiovascular disease risk communication strategies in reducing the incidence of cardiovascular disease?",
  "pmid": "24024587",
  "title": "Shared decision-making in antihypertensive therapy: a cluster randomised controlled trial",
  "title_abstract": "## title \n Shared decision-making... \n\n ## abstract \n ...",
  "label": "1"
}

Data fields (selected)

Full column dictionary: docs/DATASET.md.

Target context — meta_id, target_meta_title, target_meta_pub_date, topic_id, topic_name, research_question (the screening question, LLM-generated per target review).

Candidate paper — pmid, title, journal, authors, article_type, pub_date.

Candidate text — abstract, introduction, methods, results, discussion, conclusion, other, plus merged fields title_abstract and full_text (and *_no_tables variants with embedded tables stripped). Multi-line text is flattened to literal \n markers so record boundaries remain unambiguous. The complete Parquet values are not intended for Excel because some cells exceed Excel's 32,767-character limit; use the corresponding preview CSV for GUI review.

Label — label (0/1, see above).

Bookkeeping — *_chars (character counts) and *_llama3_tokens (Llama-3 token counts) support fair truncation/filtering across models.

Usage

This dataset is distributed through a gated public Hugging Face repository. Accept the repository's access terms, then authenticate in the environment used to load the data:

hf auth login

For non-interactive environments, provide HF_TOKEN instead. Dataset download does not require NCBI credentials, Ollama, or a GPU.

from datasets import load_dataset

data_files = {
    "train": "train_clean.parquet",
    "validation": "validation_clean.parquet",
    "test": "test_clean.parquet",
}
ds = load_dataset("pranesh-sk/meta-screening-benchmark", data_files=data_files)
row = ds["validation"][0]
print(row["research_question"])
print(row["title_abstract"])
print(row["label"])  # "1" = cited relevance proxy; "0" = non-cited hard negative

All 96 canonical fields are nullable strings by design. Cast labels, counts, or flags after loading when numeric or Boolean operations are required. Never use *_clean_excel_preview.csv for model training or evaluation: shortened text in those files is deliberately marked EXCEL PREVIEW ONLY; FULL VALUE IN SOURCE.

Do not feed label, meta_id, pmid, pmcid, target_meta_pmid, or candidate_is_evidence_synthesis to a model under evaluation — these leak the answer or the retrieval key. Full allowed/forbidden column list and a recommended prompt: docs/INFERENCE.md.

Dataset creation

Source data: PubMed (target meta-analysis search) and PMC Open Access (full-text candidate retrieval), via NCBI Entrez.

Construction pipeline, per topic:

Search PubMed for target meta-analyses/systematic reviews matching the topic query and a title keyword gate.
For each target review, fetch its cited references. Papers with retrievable PMC Open Access full text become label = 1 rows (minimum 10 usable references, or the target review is dropped).
An LLM (Llama 3, local via Ollama) generates a research question from the target review's title, then a broad PubMed search query from that question.
Every candidate query is capped at the earliest date consistent with the target review's publication record. The query retrieves topically similar papers not cited by the target review; precision-aware dates are checked again after full-text retrieval, and temporally valid candidates become label = 0 rows (target: 10x the positive count, minimum 70% of that target or the target review is dropped).
Full text is parsed from PMC JATS XML into canonical sections.

Full procedure and release contract: README.md, docs/DATASET.md, and docs/EXCEL_EXPORTS.md.

Annotations: labels are derived automatically from each target review's actual reference list—not manually annotated eligibility judgments. The task is to prioritize a small, high-recall relevance pool for a practitioner, who still makes the final inclusion decision. A deterministic QA sample (manual_debug_100_rows.csv) is included in this dataset release for spot-checking parse quality.

Considerations for use

Hard negatives, not random negatives. Label-0 papers are topically close to the target review by construction, which makes this a harder and more realistic screening task than random-negative setups — but also means recall failures are more informative than accuracy alone. Report precision/recall/F1, not just accuracy.
Recall of positives is bounded by PMC Open Access availability. A reference not available in PMC OA could not be retrieved as full text, so it's excluded from label-1 rows even if it was genuinely cited.
The pharmacology topic is not cardiology-scoped. Don't assume domain uniformity across topics when slicing results.
Single LLM in the loop for question/query generation. Research questions and search queries were generated by one local Llama 3 model; this could bias which candidates were reachable as negatives.

Licensing and citation

License: CC BY 4.0

@dataset{meta_screening_benchmark,
  title        = {Meta-Screening Benchmark: Leakage-Controlled Systematic-Review Screening for LLM Evaluation},
  author       = {Gabriel, Roy M. and Sathish Kumar, Pranesh and Galarnyk, Michael and Kittisut,
                  Nattakorn and van Assen, Marly, and De Cecco, Carlo N. and Adibi, Ali},
  year         = {2026},
  url          = {https://huggingface.co/datasets/pranesh-sk/meta-screening-benchmark}
}

Source and rebuild tooling: github.com/praneshsathishk/automated-meta-analysis-llms

Downloads last month: 90

Total file size:

1.76 GB