jsoul/geo-perturbation-grouping-test

GEO Gene-Perturbation Grouping — Held-out Test

The held-out evaluation set for the GEO gene-perturbation grouping task. The training and validation data live in the companion private repo jsoul/geo-perturbation-grouping-train, which is where the full task description, rulebook, and data-format documentation are maintained.

⚠️ Intended as a clean evaluation benchmark — avoid using it for training or model selection.

Contents

• 300 examples, one test split.
• 300 unique GEO accessions, disjoint from both the train and validation splits in the companion repo (zero leakage).
• Identical schema and system prompt to the training data.

Task & format (summary)

Given a GEO study's sample metadata, produce the valid genetic-perturbation case/control groups (methods KD / KO / OE, ≥2 controls and ≥2 cases, single wild-type HGNC target, one cell line, matched time points), or No valid groups found. Each row is a single-turn chat:

field	type	description
messages	list[{role, content}]	system (fixed rulebook) → user (study metadata) → assistant (target grouping).
metadata.accession	string	Source GEO series ID (e.g. GSE237971).

Target format:

Group <n>:
  Cell line: <cell line>
  Perturbation method: <KD | KO | OE>
  Target gene: <HGNC symbol>
  Control: <comma-separated GSM IDs>
  Case: <comma-separated GSM IDs>

See the train repo for the complete rulebook and worked examples.

Statistics

split	rows	unique accessions	No valid groups found
test	300	300	119 (39.7%)

Per-group method counts: KD 162, KO 67, OE 49. Groups per example range 0–8.

Usage

from datasets import load_dataset

test = load_dataset("jsoul/geo-perturbation-grouping-test", split="test")  # private; needs a token

# Evaluate: feed messages[0] (system) + messages[1] (user) to the model,
# compare its output against messages[2] (the reference grouping).

Licensing & provenance

Derived from publicly available NCBI GEO metadata; assistant targets are curated annotations for this project. Distributed privately for research use. See the companion train repo for full details.