BPE
Used in: GPT
Merge frequent letter pairs
WordPiece
Used in: BERT
Merge by likelihood
SentencePiece
Used in: T5, multilingual
Handles no-space languages
🔹 BPE (Byte Pair Encoding)
Used by GPT models. Merge most frequent letter pairs. Purely frequency-based.
🔹 WordPiece
Used by BERT. Merge based on likelihood improvement. Slightly smarter merging.
🔹 SentencePiece
Used in multilingual models (e.g. T5). Handles languages without spaces like Chinese.