Multilingual Pre-training is the practice of training a single model on text from many different languages simultaneously (e.g., 100 languages) — typified by mBERT and XLM-RoBERTa, allowing the model to learn universal semantic representations that align across languages.

Mechanism

• Data: Concatenate Wikipedia/CommonCrawl from 100 languages.
• Tokenizer: Use a shared sentencepiece vocabulary (typically large, e.g., 250k tokens).
• Training: Standard MLM. No explicit parallel data (translation pairs) is strictly needed, though it helps.
• Result: A model that can process input in Swahili, English, or Chinese without specifying the language.

Why It Matters

• Cross-Lingual Transfer: You can fine-tune on English labeled data and run inference on German text.
• Low-Resource Support: High-resource languages (English) help the model learn structures that transfer to low-resource languages (Swahili).
• Simplicity: One model to deploy instead of 100 separate models.

Multilingual Pre-training is the Tower of Babel solved — creating a single polyglot model that maps all languages into a shared semantic space.