Global News Retrieval Benchmark with Multihop Queries Across 10 Languages

A multilingual news retrieval benchmark containing synthetic multihop queries across 10 languages. The dataset is sourced from recent news articles and is designed for evaluating dense retrieval and text embedding models on content post-dating typical model training cutoffs. It was created by jinaai and last updated on April 30, 2026.

Use Cases

• Evaluate dense retrieval models based on multihop queries requiring information synthesis.
• Benchmark text embedding models on recent, real-world news content across multiple languages.
• Test model performance on information that likely post-dates its training data cutoff.

Strengths

• Queries are designed to require combining facts from different parts of a single article, creating a specific evaluation task.
• The corpus is sourced from recent articles, making it suitable for testing on post-training-cutoff information.
• Covers 10 languages, providing a multilingual evaluation benchmark.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment for large-scale training.

Provenance

Source

huggingface

Collection Method

Sourced from recent news articles; queries are synthetically generated.

Time Range

Recent articles (post-dating typical model training cutoffs).

Freshness

Last updated 2026-04-30 08:23:53.

Geography

Global, based on the title and multilingual scope.