PhytoBench-Paper

Evaluation framework for the In Silico Research Agent of the Phytomni multi-agent system. The benchmark probes the agent's ability to replicate the computational portion of published plant-science papers end-to-end, from data download and tool installation through analysis and figure generation.

Construction

Five representative bioinformatics papers were selected, one per plant-science domain:

scene	Paper
plant evolution	A graph-based genome and pan-genome variation of the model plant Setaria — Nature Genetics, 2023
plant genetics	Transcriptional landscape of rice roots at the single-cell resolution — Molecular Plant, 2021
plant epigenetics	Population-wide DNA methylation polymorphisms at single-nucleotide resolution in 207 cotton accessions reveal epigenomic contributions to complex traits — Cell Research, 2024
plant molecular biology	Prediction of conserved and variable heat and cold stress response in maize using cis-regulatory information — The Plant Cell, 2021
plant development	RNA 5-Methylcytosine Modification Regulates Vegetative Development Associated with H3K27 Trimethylation in Arabidopsis — Advanced Science, 2022

Each paper was chosen so that its computational analyses can be reproduced from scratch using publicly available datasets and tools. The accompanying replication rubric (not included in this dataset card) decomposes each paper into a hierarchical set of weighted gradable tasks — data download, tool installation, execution, data cleaning, visualization, and results analysis — totalling over 1,353 gradable points across the five papers.

Schema

Field	Type	Description
scene	string	Plant-science domain identifier; one of the five values listed above.
title	string	Paper title.
journal	string	Journal name.
year	int64	Publication year.
attachment_url	string	URL of the per-paper ZIP archive (input datasets + supporting files), hosted in this repository under data/.
file_size_mb	float64	Size of the attachment archive, in megabytes.

Unlike PhytoBench-Analysis, per-paper attachments are stored as loose .zip files alongside the parquet, not inlined in the row. The parquet keeps only the URL pointer and size metadata; download the archives separately when running the benchmark.

How to use

from datasets import load_dataset
from huggingface_hub import hf_hub_download

ds = load_dataset("Phytomni/PhytoBench-Paper", split="test")

for r in ds:
    print(f"[{r['scene']}] {r['title']} ({r['journal']} {r['year']}) ~{r['file_size_mb']:.1f} MB")
    # Download the paper's input archive on demand:
    zip_filename = r["attachment_url"].split("/")[-1]
    local_path = hf_hub_download(
        repo_id="Phytomni/PhytoBench-Paper",
        filename=f"data/{zip_filename}",
        repo_type="dataset",
    )
    # Then run your agent against the unpacked archive and score against the rubric.

Evaluation protocol. Replication is scored against the per-paper rubric of weighted tasks.

Citation

@article{phytomni2026,
  title  = {Phytomni: An agentic AI accelerating plant research from discovery to design},
  author = {Phytomni Team},
  year   = {2026},
  note   = {Manuscript in preparation; citation TBD until publication.}
}

Links

• Collection: https://huggingface.co/collections/Phytomni/phytobench
• Project: https://github.com/Phytomni/Phytomni
• Agent code: https://github.com/Phytomni/Phytomni-Bot

License: GPL-3.0

Version

v0.1.0