SIB-200: Multilingual Topic Classification Dataset for 205 Languages

SIB-200 is a multilingual topic classification dataset covering 205 languages and dialects. It is based on the human-translated Flores-200 corpus, with topic annotations originally provided in English for categories like science/technology, travel, and politics. The dataset was created by the mteb organization and last updated in February 2026.

Use Cases

• Benchmarking multilingual text embedding models based on topic classification performance.
• Training topic classifiers for low-resource languages based on transferred English labels.
• Analyzing cross-lingual transfer learning performance for text categorization.
• Evaluating the consistency of topic labels across human-translated parallel texts.

Strengths

• Covers 205 languages and dialects, making it one of the largest publicly available multilingual topic datasets.
• Annotations are based on human-translated text from the Flores-200 corpus.
• Topic labels are provided for seven distinct categories: science/technology, travel, politics, sports, health, entertainment, and geography.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Topic labels originate from English annotations and are transferred to other languages, which may introduce label noise.

Provenance

Source

huggingface

Collection Method

Annotations transferred from English to human-translated texts in the Flores-200 corpus.

Freshness

Last updated 2026-02-24 21:18:51; freshness should be verified.