MultiIFEval: Multilingual Instruction-Following Evaluation Dataset for 300+ Languages

300+ languages are covered by this instruction-following dataset, translated and localized from the English IFEval dataset using Gemini-3-flash-preview. The dataset was created by danish-foundation-models and was last updated on February 22, 2026. Samples include a random Wikipedia article in the target language to provide context for localization.

Use Cases

• Benchmarking instruction-following accuracy of LLMs based on the translated IFEval framework.
• Evaluating model performance across low-resource languages based on the 300+ language coverage.
• Studying the impact of cultural and linguistic localization on instruction comprehension based on the inclusion of Wikipedia articles.

Strengths

• Covers over 300 languages, providing a broad multilingual scope.
• Samples are localized with contextual Wikipedia articles, potentially improving cultural relevance.
• Based on the established IFEval benchmark for instruction-following evaluation.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Data may reflect translation or source bias inherent to the automated localization process.

Provenance

Source

Translated and localized from the English IFEval dataset.

Collection Method

Samples were translated and localized using Gemini-3-flash-preview, with the addition of random Wikipedia articles for context.

Freshness

Last updated 2026-02-22 12:13:47; freshness should be verified.