From 9c15d1f83ee2f9387126cf4892cd94f39bdbff5e Mon Sep 17 00:00:00 2001 From: aaron miller Date: Fri, 26 May 2023 11:23:42 -0700 Subject: [PATCH] add tokenizer readme w/ instructions for convert script --- gpt4all-backend/tokenizer/README.md | 14 ++++++++++++++ 1 file changed, 14 insertions(+) create mode 100644 gpt4all-backend/tokenizer/README.md diff --git a/gpt4all-backend/tokenizer/README.md b/gpt4all-backend/tokenizer/README.md new file mode 100644 index 00000000..46563a2a --- /dev/null +++ b/gpt4all-backend/tokenizer/README.md @@ -0,0 +1,14 @@ +# BPE tokenizer + +This is a C++ implementation of the encoding/decoding functions of a pretrained GPT-2 style BPE tokenizer. It is meant to be compatible with the GPT-J and MPT-7B tokenizers that were trained with HuggingFace [`tokenizers`](https://github.com/huggingface/tokenizers), and only implements the necessary functionality for those models (it is assumed that strings should always be [normalized](https://en.wikipedia.org/wiki/Unicode_equivalence) to Unicode NFC form and split with the GPT-2 "pretokenizing" [regular expression](https://github.com/karpathy/minGPT/blob/37baab71b9abea1b76ab957409a1cc2fbfba8a26/mingpt/bpe.py#L92)) + +## Converting a tokenizer file + +`scripts/gen_tokenizer_include.py` can be used to convert a huggingface `tokenizers` `tokenizer.json` file into a C++ header file: + +```bash +# get tokenizer.json +cd gpt4all-backend +wget -O /tmp/gptj-tokenizer.json https://huggingface.co/nomic-ai/gpt4all-j/raw/main/tokenizer.json +python ./scripts/gen_tokenizer_include.py /tmp/gptj-tokenizer.json gptj > ./tokenizer/gptj_tokenizer_config.h +```