wjixiang/biomnibench-da

OmicOS-BiomniBench Task Packages (Flat / Standardized)

Note: This directory is an improved, structural-flattened and format-standardized derivative of the upstream OmicOS-BiomniBench benchmark. The original release stores each task as a directory of heterogeneous files (instruction.md, task.toml, environment/, tests/rubric.txt, ...). This package reformats every task into the single-file, schema-versioned task-package layout consumed by our downstream agent harness, so that all 50 tasks share one uniform shape and can be loaded by generic tooling without task-specific parsing.

The layout and portable data formats have been normalized. Excel workbooks were flattened to one TSV per sheet, MACS2 .xls peak tables were converted to TSV, unsupported .RData matrices were converted to Parquet, and duplicate CSV intermediates from conversion were removed. No .xlsx, .xls, or .RData files or references remain. The source is noncommercial benchmark data — follow the upstream license and dataset terms.

What changed vs. the original

Aspect	Upstream OmicOS-BiomniBench	This package (flat)
Per-task layout	Multi-file directory (instruction.md, task.toml, environment/, tests/...)	Single task.json + data/ directory
Instruction	Free-form instruction.md	Embedded in task.json under instruction
Harbor config	Separate task.toml	Embedded in task.json under metadata.harbor_task
Expert rubric	Separate tests/rubric.txt	Embedded in task.json under grading.rubric
Data reference	environment/ copied or referenced per task	data array of {name, path, media_type, role} links to the shared environment/
Schema	Implicit, TOML + Markdown	Explicit schema_version: 1 JSON; top-level id, category, subtask, data, instruction, grading, metadata
Loader	Task-specific parsing of .md/.toml/.txt	One generic JSON loader for every task
Tabular formats	Mixed Excel, R serialization, CSV, TSV	No .xlsx / .xls / .RData; converted sheets use TSV, converted matrices use Parquet, source CSV files are retained

The initial structural flattening is driven by convert_omicos_biomnibench_to_task_packages.py; the subsequent format normalization updates task instructions, rubrics, Dockerfiles, filenames, and file references together. The current package contains 110 CSV files, 69 TSV files, 2 Parquet files, and 2 Matrix Market files. Matrix Market files were intentionally retained because they are directly supported by the target runtime. One instruction/Dockerfile reference to an RData marker file absent from the released source snapshot was removed rather than replaced with a fabricated input.

Layout

This directory contains 50 BiomniBench-DA task packages converted from the source snapshot. Each task is one subdirectory under biomnibench_da/:

OmicOS-BiomniBench-task-packages-flat/
└── biomnibench_da/
    ├── da-1-3/
    │   ├── task.json     # unified task-package document (schema v1)
    │   └── data/         # flattened task data plus Dockerfile
    ├── da-1-4/
    │   ├── task.json
    │   └── data/
    └── ...               # 50 tasks in total

task.json schema (v1)

Top-level fields:

• schema_version — integer schema version (currently 1).
• id — canonical task identifier, e.g. da-1-3.
• category — task category (e.g. oncology).
• subtask — optional sub-classification.
• instruction — full task prompt in Markdown (originally instruction.md).
• data — list of workspace assets. Each entry has name, path, media_type, and role (typically workspace) and points at the task's flattened data directory.
• grading — answer, answer_type (e.g. rubric), method (e.g. rubric), pass_threshold, and the full rubric text (originally tests/rubric.txt).
• metadata — task-level metadata: difficulty, source, and the full Harbor harbor_task block (agent, environment, artifacts, …) originally from task.toml.

Usage

import json
from pathlib import Path

root = Path("OmicOS-BiomniBench-task-packages-flat/biomnibench_da")
for task_dir in sorted(root.iterdir()):
    pkg = json.loads((task_dir / "task.json").read_text())
    assert pkg["schema_version"] == 1
    print(pkg["id"], pkg["category"], pkg["metadata"]["difficulty"])

Downstream agents and graders should consume task.json directly and read data files from the linked data/ directory; they should not re-parse the upstream instruction.md / task.toml / rubric.txt.

Provenance & licensing

• Upstream dataset: phylobio/BiomniBench-DA (CC-BY-4.0).
• Source snapshot: OmicOS-BiomniBench/ (see its LICENSE).
• This package introduces no new primary observations; it is a structural and format-normalization layer over the upstream release. The source is noncommercial benchmark data — follow the upstream license and dataset terms.