LLM Learning Hub

workspace/llm-course/home

SentencePiece

SentencePiece is a language-independent subword tokenizer that treats the input as a raw byte stream, so it works on any language without pre-tokenization on whitespace. It supports BPE and Unigram modes.

What is SentencePiece?

SentencePiece is a language-independent subword tokenizer that treats the input as a raw byte stream, so it works on any language without pre-tokenization on whitespace. It supports BPE and Unigram modes.

SentencePiece is a language-independent subword tokenizer that treats the input as a raw byte stream, so it works on any language without pre-tokenization on whitespace. It supports BPE and Unigram modes.

Where is it used?

SentencePiece is used by Llama 2/3, T5, ALBERT, and multilingual models like XLM-R; its whitespace-agnostic design is crucial for languages without spaces (Chinese, Japanese).

How to build it

Install `sentencepiece`, train with `spm.SentencePieceTrainer.train('--input=corpus.txt --model_prefix=m --vocab_size=32000'`, and encode/decode with `spm.SentencePieceProcessor`.