Datasetpublic

PhilipQuirke/tinyScienceQA

tinyScienceQA Pathfinder-generated 100-example tiny subset of the text-only ScienceQA test split. This is intended for fast smoke tests and default small-sample Pathfinder runs, not as an official ScienceQA benchmark replacement. Construction source dataset: tasksource/ScienceQA_text_only source split: test source rows: 2224 tiny rows: 100 seed: 20260511 selection strategy: subject_answer_quota_topic_coverage_stable_hash The selector preserves… See the full description on the dataset page: https://huggingface.co/datasets/PhilipQuirke/tinyScienceQA.

sourceHugging Facecc-by-sa-4.0updated 5mo agoView on Hugging Face

0likes21downloads

Dataset Card

tinyScienceQA

Pathfinder-generated 100-example tiny subset of the text-only ScienceQA test split. This is intended for fast smoke tests and default small-sample Pathfinder runs, not as an official ScienceQA benchmark replacement.

Construction

—source dataset: tasksource/ScienceQA_text_only
—source split: test
—source rows: 2224
—tiny rows: 100
—seed: 20260511
—selection strategy: subject_answer_quota_topic_coverage_stable_hash

The selector preserves subject/answer-cell proportions with exact integer quotas, ensures topic coverage where quota constraints permit, and fills the remaining slots by a deterministic stable hash. The resulting subset covers 25 of 25 topics.

Counts

Subject counts:

json

{
  "language science": 47,
  "natural science": 46,
  "social science": 7
}

Answer counts:

json

{
  "0": 44,
  "1": 43,
  "2": 9,
  "3": 4
}

The corresponding Pathfinder benchmark id is scienceqa, and the default tiny dataset id is PhilipQuirke/tinyScienceQA.