license: mit
language:
- en
tags:
- tokeniser
pretty_name: tokeniser on val, test set of Slim Pajama
size_categories:
- 1B<n<10B
Tokenizer
Imp Links: PyPI Main Library (tokeniser-py) | PyPI Lite Library (tokeniser-py-lite) | Main Library GitHub (tokeniser-py) | Lite Library GitHub (tokeniser-py-lite) | Demo (HF Spaces) | Complete repo (chunked) - GitHub | Imp Files Github
This is a tokeniser created on a custom-written algorithm on a huge vocabulary of ~1B tokens. These tokens are given in the files (such that they are <2GB each, making them trackable by Git LFS). The text corpus is from the SlimPajama dataset by cerebras and consists of the whole text and validation corpus.
The final tokeniser is available in two versions (0.5B version - Val. data only and 1B version - Val data + Test data, created using the same algo).
The files includes the token counts, the text corpus used, individual lines/paras from SlimPajama as a list JSON, ordered tokeniser with token ids (in order of their counts), unordered tokeniser with token ids.
The tokeniser contains 131,072 tokens.
To do:
- Write custom code for the final tokenisation part (to break text into tokens)
- Create a python library for using the tokeniser
- Put the files on GitHub for a general overview
- Release the experimentation notebook and the tokenisation code (as part of the library)
- Make a writeup to explain the algo used
Note: algo used is no industry standard algo like Byte Pair encoding, infact i have not studied BPE yet, i wanted to create a tokeniser from scratch first without any idea of what is currently used, and then compare the two, so a lot of what i implement may have similarities to that in industry but may not provide some performance improvement
I am storing it on HF, not on GITHUB, because of storage issues, and due to easy availability for the AI community.