* fix truncate * s/SentencePieceModel/SentencePiece/ * bert * wordpiece * refactor pooling * more tokenizers * normalize embeddings