Datasets:

domsoos
/

replicatorbench

This repository contains the data-only portion of the Center for Open Science (COS) LLM Benchmarking Project. The dataset supports benchmarking LLM agents on core parts of the scientific research lifecycle—especially replication—including:

Information extraction from scientific papers into structured JSON
Research design and analysis planning
(Optional) execution support using provided replication datasets and code
Scientific interpretation using human reference materials and expected outputs

Each numbered folder corresponds to one study instance in the benchmark.

Dataset contents (per study)

Each study folder typically contains:

original_paper.pdf
The published paper used as the primary input.
initial_details.txt
Brief notes to orient the replication attempt (e.g., key outcomes, hints, pointers).
replication_data/
Data and scripts required to reproduce analyses (common formats: .csv, .dta, .rds, .R, .do, etc.).
human_preregistration.(pdf|docx)
Human-created preregistration describing the replication plan.
human_report.(pdf|docx)
Human-created replication report describing analyses and findings.
expected_post_registration*.json
Expert-annotated ground truth structured outputs used for evaluation.
- expected_post_registration.json is the primary reference.
- expected_post_registration_2.json, _3.json, etc. are acceptable alternative variants where applicable.

Some studies include multiple acceptable ground-truth variants to capture permissible differences in annotation or representation.

Repository structure

At the dataset root, folders like 1/, 2/, 10/, 11/, etc. are study IDs.

Example:

text
.
├── 1/
│   ├── expected_post_registration.json
│   ├── expected_post_registration_2.json
│   ├── human_preregistration.pdf
│   ├── human_report.pdf
│   ├── initial_details.txt
│   ├── original_paper.pdf
│   └── replication_data/
│       ├── <data files>
│       └── <analysis scripts>

Intended uses

This dataset is intended for:

Benchmarking LLM agents that extract structured study metadata from papers
Evaluating LLM systems that generate replication plans and analysis specifications
Comparing model outputs against expert-annotated expected JSON and human reference docs

Not intended for

Clinical or other high-stakes decision-making
Producing definitive judgments about the original papers
Training models to reproduce copyrighted texts verbatim

Quickstart (local)

Iterate over studies and load ground truth

python
from pathlib import Path
import json

root = Path(".")
study_dirs = sorted(
    [p for p in root.iterdir() if p.is_dir() and p.name.isdigit()],
    key=lambda p: int(p.name)
)

for study in study_dirs:
    gt = study / "expected_post_registration.json"
    if gt.exists():
        data = json.loads(gt.read_text(encoding="utf-8"))
        print(study.name, "ground truth keys:", list(data.keys())[:10])

Using with the main pipeline repository (recommended)

If you are using the LLM Benchmarking Project codebase, point the pipeline/evaluators at a given study directory:

bash
make evaluate-extract STUDY=/path/to/llm-benchmarking-data/1

The expected JSON format is defined by the main repository’s templates/schemas. Use those schemas to validate or format model outputs.

Notes on multiple expected JSON variants

Some studies include expected_post_registration_2.json, expected_post_registration_3.json, etc. This is intentional:

Some fields allow multiple equivalent representations
Annotation may vary slightly without changing meaning
Evaluators may accept any variant depending on scoring rules

If you implement your own scorer, consider:

Exact matching for strictly defined fields
More tolerant matching for lists, notes, or fields with legitimate paraphrase/format variation

File formats

You may encounter:

R artifacts: .R, .rds
Stata artifacts: .do, .dta
CSV/tabular data: .csv
Documents: .pdf, .docx
Structured evaluation targets: .json

Reproducing analyses may require R and/or Stata depending on the study.

Licensing, copyright, and redistribution (important)

This repository is released under Apache 2.0 for COS-authored materials and annotations (for example: benchmark scaffolding, expected JSON outputs, and other COS-created files).

However, some contents may be third-party materials, including (but not limited to):

original_paper.pdf (publisher copyright may apply)
replication_data/ (may have its own license/terms from the original authors)
external scripts or datasets included for replication

You are responsible for ensuring you have the right to redistribute third-party files publicly (e.g., GitHub / Hugging Face).

Common options if redistribution is restricted:

Remove third-party PDFs and provide DOI/URL references instead
Keep restricted files in a private location and publish only COS-authored annotations
Add per-study LICENSE / NOTICE files inside each study folder where terms are known

Large files (Git LFS recommendation)

If hosting on GitHub, consider Git LFS for PDFs and large datasets:

bash
git lfs install
git lfs track "*.pdf" "*.dta" "*.rds"
git add .gitattributes

Citation

If you use this dataset in academic work, please cite it as:

bibtex
@dataset{cos_llm_benchmarking_data_2026,
  author    = {Center for Open Science},
  title     = {LLM Benchmarking Project: Scientific Replication Benchmark Data},
  year      = {2026},
  publisher = {Center for Open Science},
  note      = {Benchmark dataset for evaluating LLM agents on scientific replication tasks}
}

Acknowledgements

This project is funded by Coefficient Giving as part of its “Benchmarking LLM Agents on Consequential Real-World Tasks” program. We thank the annotators who contributed to the ground-truth post-registrations for the extraction stage.

Contact

For questions about this dataset:

Shakhlo Nematova
Research Scientist, Center for Open Science
shakhlo@cos.io