Tokenisation sous-baatu

A word-only vocabulary allocates a separate entry to each known word form and can require an unknown-word treatment for unseen forms. Subword methods instead compose text from reusable pieces. This can make a finite vocabulary useful for many rare or novel combinations, but coverage still depends on the alphabet, normalization, and any fallback configured. A piece can be a complete word or a single character; it need not always lie strictly between those sizes. BPE, WordPiece, and Unigram have different training and segmentation mechanisms. BPE grows larger symbols through pair merges. The original BERT uses WordPiece, with its published tokenizer marking word-internal continuation pieces using the prefix ##. That convention describes the piece vocabulary; it does not mean every visible suffix must become a separate token. If related word forms share an available piece, a model can reuse its representation, but the tokenizer is not required to discover linguistically perfect stems or morphemes. Unigram vocabulary training takes another approach, starting with many candidate pieces and pruning them according to their effect on the likelihood objective. SentencePiece should be understood separately from a single algorithm: it is a tokenizer and detokenizer library supporting BPE and Unigram. It can train directly from raw sentences without requiring language-specific word pre-tokenization. Whitespace is represented through a visible meta-symbol in its pieces, helping reconstruct normalized text. As with other tokenizers, vocabulary pieces map to integer IDs before model processing. An example split is illustrative unless a particular model file and encoding setup have been identified.