PhytoBench-Data

Evaluation benchmark for the Data Agent of the Phytomni multi-agent system. The benchmark probes natural-language-to-SQL translation over a crop multi-omics database, with an emphasis on cross-table generalization.

Construction

1,020 Q&A pairs spanning representative application scenarios across 65 plant species. Questions are designed so that correct answers require multi-table associations and stress the agent's ability to plan complex cross-table queries rather than retrieve from a single column.

Schema

Field	Type	Description
scene	string	Application scenario identifier; selects the table family or query type.
query	string	Natural-language question posed to the agent.
answer	string	Reference answer (executable SQL or the expected query result, depending on the scenario).

How to use

from datasets import load_dataset

ds = load_dataset("Phytomni/PhytoBench-Data", split="test")
sample = ds[0]
print(sample["scene"], "::", sample["query"])
print("ref:", sample["answer"])

Evaluation protocol. Translate each query into SQL with your agent, execute the SQL against the underlying multi-omics tables, and compare the result to answer. Score by exact match or execution accuracy depending on the scenario family.

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