jsoul/geo-perturbation-grouping-production

GEO Gene-Perturbation Grouping — Production / Inference Inputs

The production inference corpus for the GEO gene-perturbation grouping task: a large set of NCBI GEO studies formatted as model prompts, to be run through a fine-tuned model (e.g. `jsoul/geo-perturbation-grouping-qwen3-30b-a3b` or `jsoul/geo-perturbation-grouping-qwen3-4b`) to generate case/control perturbation groupings at scale.

⚠️ Unlabeled — inputs only. Every row contains just a system + user message and no `assistant` target. This is not a training or evaluation dataset; it is the set of prompts to run inference on.

Contents

• —6,490 examples, one split named `validation`.
• —6,490 unique GEO accessions.
• —Each row: messages = [system, user] (the fixed task rulebook + one study's sample metadata) and metadata.accession (the GEO series ID, e.g. GSE147159).
• —The system prompt is identical to the one used in the training data.

Note on the split name: the single split is named validation (rather than something like production) because downstream tooling depends on that exact name. It does not imply this is an evaluation set — see the overlap note below.

Data format

Expected model output (generated at inference time, not included here):

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>

or No valid groups found.

⚠️ Overlap with the labeled splits — read before evaluating

This corpus is a broad production sweep of GEO and overlaps the training universe by GEO accession:

Because of the train/validation overlap, do not treat this set as a held-out benchmark. For clean evaluation use `jsoul/geo-perturbation-grouping-test`, which is disjoint from train, validation, and this corpus.

Usage

from datasets import load_dataset

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

row = ds[0]
system = row["messages"][0]["content"]
user   = row["messages"][1]["content"]
# feed [system, user] through the model to generate the grouping for row["metadata"]["accession"]

Licensing & provenance

Derived from publicly available NCBI GEO metadata. Distributed privately for research use. See the companion train / test dataset cards for full task details and the rulebook.