SearchTome

SearchTome is a structure-aware retrieval benchmark spanning six domains. Each domain provides a corpus drawn from a single open-access textbook and synthetically generated query/answer pairs split into train / dev / test. It accompanies the STAIR (STructure-Aware Information Retriever) work.

Dataset structure

Two config types per domain:

• <domain>_corpus — one corpus split, records: {title, content}. title is the document identifier.
• <domain>_queries — train / dev / test splits, records: {query_id, query, gold_doc_title, answer}. query_id is a 0-based index within each file. gold_doc_title is the relevant document and matches a corpus title exactly (verified at build time). answer is the synthetic answer.

Domains: education, finance, law, medicine, natural_sciences, social_sciences.

Corpus titles are unique within each domain (verified at build time), so each query's gold_doc_title resolves to exactly one document.

Sizes

Domain	Corpus docs	Train	Dev	Test
education	8	308	86	437
finance	86	940	180	1214
law	47	591	152	820
medicine	121	1473	346	1996
natural_sciences	34	435	94	579
social_sciences	117	658	198	959

Loading

from datasets import load_dataset

# corpus for one domain
corpus = load_dataset("meghanadh/SearchTome", "medicine_corpus", split="corpus")

# question split for one domain
test = load_dataset("meghanadh/SearchTome", "medicine_queries", split="test")

# map a query to its gold document
title_to_doc = {row["title"]: row["content"] for row in corpus}
gold = title_to_doc[test[0]["gold_doc_title"]]

Source corpora & attribution

Each corpus is derived from an open-access textbook. Attribution is required under both CC BY and CC BY-SA. Two sources (finance, social_sciences) are ShareAlike, which is why the whole release is distributed under Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0).

Domain	Source textbook	Author(s)	Publisher	Original license
education	Action Research	J. Spencer Clark; Suzanne Porath; Julie Thiele; Morgan Jobe	New Prairie Press, Kansas State University (2020)	CC BY 4.0
finance	Principles of Financial Accounting	Christine Jonick	University of North Georgia Press (2017)	CC BY-SA 4.0
law	Introduction to Intellectual Property	OpenStax	OpenStax, Rice University (2021)	CC BY 4.0
medicine	Nursing Assistant	Open RN, Chippewa Valley Technical College	WisTech Open (2023)	CC BY 4.0
natural_sciences	Introduction to Soil Science	Amber Anderson; Rivka Fidel	Iowa State University Digital Press (2023)	CC BY 4.0
social_sciences	Foundations of Aural Skills	Timothy Chenette	Utah Education Network (UEN) (2022)	CC BY-SA 4.0

License

The whole release — corpora and synthetically generated questions — is licensed under Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0) (https://creativecommons.org/licenses/by-sa/4.0/).

CC BY-SA is the only single license compatible with all six source textbooks: the four CC BY sources may be redistributed under CC BY-SA, and the two CC BY-SA sources require it. You must retain the attributions above and share derivatives under the same license.

Citation

@misc{searchtome,
  title        = {SearchTome: A Structure-Aware Multi-Domain Retrieval Benchmark},
  author       = {Kumar, Vineet and Pulivarthi, Meghanadh and Kumar, Vishwajeet and Sen, Jaydeep and Bhat, Riyaz Ahmad and Joshi, Sachindra},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/meghanadh/SearchTome}},
  note         = {Corpora derived from open textbooks; see attribution table}
}