CoolFace
Datasetpublic

sastpg/aaai27-hotpotqa-fullwiki-original

HotpotQA FullWiki frozen original data Private reproducibility snapshot for the AAAI 2027 experiments. This repository stores the exact Hugging Face hotpotqa/hotpot_qa, fullwiki parquet files used to construct our train, calibration, and official evaluation task identities. It does not contain generated trajectories, Writer SFT/RL examples, or model-selected subsets. Files and rows fullwiki/train-00000-of-00002.parquet: 45,224 rows.… See the full description on the dataset page: https://huggingface.co/datasets/sastpg/aaai27-hotpotqa-fullwiki-original.

sourceHugging Facecc-by-sa-4.0updated 2mo agoView on Hugging Face
0likes59downloads
Dataset Card

HotpotQA FullWiki frozen original data

Private reproducibility snapshot for the AAAI 2027 experiments.

This repository stores the exact Hugging Face hotpotqa/hotpot_qa, fullwiki parquet files used to construct our train, calibration, and official evaluation task identities. It does not contain generated trajectories, Writer SFT/RL examples, or model-selected subsets.

Files and rows

  • —fullwiki/train-00000-of-00002.parquet: 45,224 rows.
  • —fullwiki/train-00001-of-00002.parquet: 45,223 rows.
  • —fullwiki/validation-00000-of-00001.parquet: 7,405 rows.
  • —fullwiki/test-00000-of-00001.parquet: 7,405 rows.
  • —SHA256SUMS: exact file digests.

Our frozen protocol draws Writer construction/RL questions from the 90,447-row training split and draws calibration100/official500 from the validation split. Each question permits at most four search calls and each call returns top-3 documents.

Retrieval corpus

The question dataset and retrieval corpus are separate. Retrieval uses the public Wiki-18 corpus:

text
URL: https://huggingface.co/datasets/PeterJinGo/wiki-18-corpus/resolve/main/wiki-18.jsonl.gz
compressed bytes: 5123307260
compressed SHA-256: 7abd929223399cd63c52b499f289bf4f9039be1e9f8c43e1cb3938305b2317db
uncompressed rows: 21015325
uncompressed SHA-256: 85a787a692e73fdd657e86411bfd1ac810ed193b1eea4a478de449d4871c06b9

The BM25 index is a derived artifact and should be rebuilt from that exact corpus. It is not duplicated in this repository.

Upstream

  • —Dataset: https://huggingface.co/datasets/hotpotqa/hotpot_qa
  • —Paper: https://arxiv.org/abs/1809.09600
  • —License: CC BY-SA 4.0.