AutoMat is a benchmark of reproducibility claims from computational materials-science papers. Access is gated to keep the leaderboard meaningful and to respect upstream paper licensing. Please briefly describe how you intend to use the dataset.

AutoMat

AutoMat is a benchmark of reproducibility claims drawn from the computational materials-science literature. Each claim packages a published scientific statement together with the inputs, papers, and reference outputs an autonomous agent (or human) would need to attempt a faithful reproduction.

Layout

Each claim lives under claims/AUTOMAT-XXXX/ with three subtrees:

Path	Contents
`meta/provenance.json`	Authoritative metadata (claim ID, paper, DOI, author).
`meta/claim.md`	Human-readable claim statement and reproduction instructions.
`agent_view/`	What an agent is allowed to see: `claim.txt`, the paper (open-access PDF, or a `.placeholder` for paywalled papers — see Licensing), input data, custom code.
`reference/`	Ground-truth reproduction: the `reproduction.txt` steps and expected result, plus a reference implementation.

Coverage

A small number of claims are withheld from the current release at their authors' request, because the underlying paper is still pending official publication. These will be added in a later revision once the corresponding papers are public. The published manifest.parquet lists exactly the claims included in this release, so any consumer that drives off the manifest stays in sync automatically.

Using the dataset

This is a gated dataset, so you must be logged into your Hugging Face account (and have been granted access) before downloading. Verify with:

hf auth whoami

If that reports you are not logged in, run:

hf auth login

Once you are logged in, you can download the dataset:

from huggingface_hub import snapshot_download
import pandas as pd

local = snapshot_download(repo_id="jhu-clsp/AutoMat", repo_type="dataset")
manifest = pd.read_parquet(f"{local}/manifest.parquet")
print(manifest.head())

To fetch a single claim only:

from huggingface_hub import snapshot_download
snapshot_download(
    repo_id="jhu-clsp/AutoMat",
    repo_type="dataset",
    allow_patterns=["manifest.parquet", "claims/AUTOMAT-0003*"],
)

The companion harness at https://github.com/JHU-CLSP/AutoMat consumes this layout directly.

Evaluation protocol

For each claim, an agent receives the contents of agent_view/ and the free-form claim text. A successful reproduction is one whose outputs match the expected result described in reference/reproduction.txt (per the comparison rules described in the harness harness/evaluators/). Reference implementations under reference/impl/ are provided for context and to debug failed runs — they are not intended as the canonical reproduction path.

Licensing notes

The metadata, claim statements, and reference outputs in this dataset are released under CC-BY-4.0.
Source papers are not redistributed for paywalled venues. To respect publisher copyright, paper PDFs from subscription publishers are not included: each such claim's agent_view/paper/ contains a .placeholder instead of the PDF. Obtain the paper yourself through your institutional access using the DOI in meta/provenance.json (the identifier field). The companion harness ships a paper_references.tsv mapping every claim to its paper title and DOI.
Only open-access (e.g. arXiv) PDFs are bundled under agent_view/paper/; these remain the property of their authors and are included for reproducibility research under the gated-access terms above.

Citation

@article{huang2026automat,
  title     = {Can Coding Agents Reproduce Findings in
                Computational Materials Science?},
  author    = {Huang, Ziyang and Cao, Yi and
                Shargh, Ali K. and Luo, Jing and
                Mei, Ruidong and Zaki, Mohd and
                Liu, Zhan and Bunstine, Wyatt and
                Jurayj, William and Goswami, Somdatta and
                McQueen, Tyrel and Shields, Michael and
                El-Awady, Jaafar and Clancy, Paulette and
                Van Durme, Benjamin and Andrews, Nicholas and 
                Walden, William and Khashabi, Daniel},
  journal = {arXiv preprint arXiv:2605.00803},
  year    = {2026},
}

Changelog

v1.0.0 — initial release.

Downloads last month: 25

Total file size:

36.4 GB

Paper for sheepy928/AutoMat

Can Coding Agents Reproduce Findings in Computational Materials Science?

Paper • 2605.00803 • Published May 1