openwebtext / README.md
sytelus's picture
add Python code
3f70adb verified
metadata
license: cc0-1.0

OpenWebText

This is the arrow format of OpenWebText dataset so that it can easily be used with HuggingFace APIs without any need for pre-processing.

Usage

This is the dataset that can be used with Andrej Karpathy's NanoGPT as well as @sytelus NanuGPU to reproduce GPT-2 series of model.

For more info on reproducing GPT-2 with NanuGPT code, see here.

Simple dataset loading:

from datasets import load_dataset
dataset = load_dataset("sytelus/openwebtext")

Statistics

Number of documents: 8013769
tokens: 9040017095 (using tiktoken GPT2 tokenizer, vocab size 50257)
Size on disk: 39,770,909,229 bytes