GutenBench

Queries and relevance judgments (qrels) for GutenBench, a diagnostic information-retrieval benchmark that stratifies queries by the retrieval capability they demand. The corpus is the Protestant biblical canon segmented into 2,813 fixed passages (pericopes), which makes the benchmark compact, stable, and available in hundreds of translations for cross-lingual evaluation.

The current release contains 2,170 fully adjudicated English queries with 10,753 graded relevance judgments.

Taxonomy levels

Each query is assigned the lowest retrieval capability required to recover its supporting evidence (see the paper for the full definitions):

Level	Capability	Queries
L1	Lexical matching	1,446
L2	Semantic matching	407
L3	Explicit multi-document aggregation	214
L4	Latent conceptual bridging	103

Structure

• topics — id, text, level (1–4, the taxonomy level above)
• qrels — topic_id, doc_id, relevance (graded: 0 = not relevant, 1 = partially relevant, 2 = definitely relevant)
• documents — doc_id, start, end (passage boundaries as OSIS references, e.g. GEN.1.1)

The level of a qrel is recoverable by joining qrels.topic_id on topics.id.

The corpus text itself is not distributed. The documents config provides only passage boundaries; pair them with a legally obtained Bible translation. For baselines we recommend widely available public-domain translations such as the King James Version or the World English Bible.

Usage

from datasets import load_dataset

topics = load_dataset("contemmcm/gutenbench", "topics", split="test")
qrels = load_dataset("contemmcm/gutenbench", "qrels", split="test")

level_4 = topics.filter(lambda q: q["level"] == 4)

GutenBench is a living dataset; releases are tagged, so pin one for reproducibility:

load_dataset("contemmcm/gutenbench", "qrels", split="test", revision="v2026")

License

CC BY 4.0

Citation

@inproceedings{monteiro-etal-2026-gutenbench,
  title     = {{GutenBench}: A Taxonomy-Based Dataset for Evaluation of
               Information Retrieval Models},
  author    = {Monteiro, M{\'a}rcio and Senkin, Denys and
               Ravichander, Abhilasha and Kloft, Marius and Fellenz, Sophie},
  booktitle = {Proceedings of the 2026 Conference on Empirical Methods in
               Natural Language Processing (EMNLP)},
  year      = {2026}
}