xnli

**XNLI (Cross-lingual Natural Language Inference)** is a **multilingual NLI benchmark spanning 15 languages** — testing whether models can perform natural language inference across languages, evaluating cross-lingual transfer and multilingual understanding. **What Is XNLI?** - **Type**: Cross-lingual NLI evaluation benchmark. - **Languages**: 15 languages including English, French, German, Chinese, Arabic, etc. - **Source**: Human translations of MultiNLI development/test sets. - **Task**: Entailment/contradiction/neutral classification across languages. - **Purpose**: Evaluate multilingual and cross-lingual models. **Why XNLI Matters** - **Multilingual**: Standard benchmark for multilingual models. - **Cross-lingual Transfer**: Test zero-shot transfer to new languages. - **Coverage**: 15 diverse languages (different families, scripts). - **Quality**: Professional human translations. - **Standard**: Used for mBERT, XLM-R, multilingual GPT evaluation. **Languages Covered** English, French, Spanish, German, Greek, Bulgarian, Russian, Turkish, Arabic, Vietnamese, Thai, Chinese, Hindi, Swahili, Urdu. **Evaluation Scenarios** - **Translate-Train**: Train on translated data. - **Translate-Test**: Translate test to English, use English model. - **Zero-Shot**: Train English only, test all languages. XNLI is the **gold standard for multilingual NLU** — testing cross-lingual generalization.