jhu-clsp/narrative-bench

NarrativeBench

NarrativeBench evaluates the information seeking capabilities models need to help people navigate diverse narratives across languages. The dataset accompanies The Narrative Gap: Can LLMs Help Us Navigate Diverse Narratives Across Languages?

GitHub · Dataset · Croissant metadata

Contents

Table	Rows	Contents
documents	1,333	News articles in 23 publication languages across 23 contested topics
queries	460	Reddit-grounded questions and translations across 24 query languages
evaluation	4,800	Document contexts, queries, instructions and scoring targets
responses	72,000	Recorded answers, language labels and scores from 15 model configurations

DATA.md describes the columns and how the tables join. models.json provides generation settings, and languages.json maps language codes to names.

Construction and tasks

Articles were gathered through localized searches and grouped by source perspective and publication language. The query set combines 283 directly sourced Reddit questions with 177 topic-specific adaptations, comprising 230 neutral and 230 confirmatory queries.

Evaluation covers Narrative Inclusion (NI), Narrative Exclusion (NE), Missing Narrative Identification (MNI), Baseline Prompt (BP), and Diverse Prompt (DP). Each selected narrative contributes three articles to a case. Models answer in the query language and cite the supplied document IDs. NI and NE use narrative F1; MNI, BP and DP use narrative recall. Scores range from zero to one, and automatic language checks use Lingua and fastText; low-confidence detections pass the language check, while responses confidently identified as outside the query language receive zero.

Load and evaluate

from datasets import load_dataset

articles = load_dataset("jhu-clsp/narrative-bench", "documents", split="train")
queries = load_dataset("jhu-clsp/narrative-bench", "queries", split="train")
cases = load_dataset("jhu-clsp/narrative-bench", "evaluation", split="test")
responses = load_dataset("jhu-clsp/narrative-bench", "responses", split="test")

Install the datasets package to use this example. The evaluation guide explains how to run the benchmark, score new answers and reproduce the recorded scores.

Citation

@misc{sharma2026narrativebench,
  title = {The Narrative Gap: Can LLMs Help Us Navigate Diverse Narratives Across Languages?},
  author = {Sharma, Nikhil and Marchisio, Kelly and Murray, Kenton and Xiao, Ziang},
  year = {2026}
}

License

Dataset annotation and curation metadata use CC BY-NC-SA 4.0. Source-article rights remain with their publishers; URLs and attribution accompany the articles. See LICENSE_NOTICE.md.