Datasets:

Phytomni
/

PhytoBench-Analysis

Parquet error: Scan size limit exceeded: attempted to read 1224391269 bytes, limit is 300000000 bytes Make sure that 1. the Parquet files contain a page index to enable random access without loading entire row groups2. otherwise use smaller row-group sizes when serializing the Parquet files

Error code:   TooBigContentError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

PhytoBench-Analysis

Evaluation benchmark for the Analyst Agent of the Phytomni multi-agent system. The benchmark covers 10 bioinformatics domains with five tasks per domain, totalling 50 end-to-end analysis scenarios. Each task ships with a binary attachment containing the input dataset(s) the agent needs in order to plan and execute the analysis.

Construction

For each of the 10 domains, five datasets of similar task type were assembled. The 10 domains are: transcriptomics, proteomics, metabolomics, evolutionary analysis, epigenetics, GWAS, genomic selection, variant calling, single-cell analysis, and spatial analysis. Tasks were chosen to capture realistic bioinformatics workflows that require tool selection, parameterization, and multi-step execution rather than single-call answers.

The scene field uses the following identifiers:

`scene` value	Domain
`transcriptome`	transcriptomics
`proteomics`	proteomics
`metabolome`	metabolomics
`evolution`	evolutionary analysis
`epigenetic`	epigenetics
`genome-gwas`	GWAS
`genome-gs`	genomic selection
`genome-callsnp`	variant calling
`single_cell`	single-cell analysis
`spatial`	spatial analysis

Schema

Field	Type	Description
`scene`	string	Domain identifier; one of the 10 values above (5 tasks per domain).
`query`	string	Natural-language analysis task posed to the agent.
`attachment`	binary	Inline ZIP archive containing the input dataset(s) for the task.

The full release is ~1.4 GB on disk; binary attachments dominate the size.

How to use

import io, zipfile
from datasets import load_dataset

ds = load_dataset("Phytomni/PhytoBench-Analysis", split="test")
sample = ds[0]
print(sample["scene"], "::", sample["query"][:160])
with zipfile.ZipFile(io.BytesIO(sample["attachment"])) as z:
    print("attachment files:", z.namelist())

Evaluation protocol. Each scenario is scored along four dimensions: (i) planning quality, (ii) tool selection, (iii) parameter setting, and (iv) completion rate. All task plans are normalized to 4–5 steps. Planning quality is assessed from the logical correctness and progression of the subtasks. Tool selection and parameter setting are assessed for correctness at each planned step. Completion rate is calculated as the proportion of the standardized pipeline completed. Each scenario is repeated five times, and completion rate is averaged across repetitions.

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.}
}