Pretokenized GitHub Code Dataset

Dataset Description

This is a pretokenized version of the Python files of the GitHub Code dataset, that consists of 115M code files from GitHub in 32 programming languages. We tokenized the dataset using BPE Tokenizer trained on code, available in this repo. Having a pretokenized dataset can speed up the training loop by not having to tokenize data at each batch call. We also include ratio_char_token which gives the ratio between the number of characters in a file and the number of tokens we get after tokenization, this ratio can be a good filter to detect outlier files.

How to use it

To avoid downloading the whole dataset, you can make use of the streaming API of datasets. You can load and iterate through the dataset with the following two lines of code:

from datasets import load_dataset

ds = load_dataset("loubnabnl/tokenized-github-code-python", streaming=True, split="train")
print(next(iter(ds)))
#OUTPUT:
{'input_ids': [504, 1639, 492,...,199, 504, 1639],
 'ratio_char_token': 3.560888252148997
}