C4#
- class streaming.text.C4(tokenizer_name, max_seq_len, group_method, local, remote=None, split=None, shuffle=True, prefetch=100000, keep_zip=None, retry=2, timeout=60, hash=None, batch_size=None)[source]#
Implementation of the C4 (Colossal Cleaned Common Crawl) dataset using streaming Dataset.
- Parameters
tokenizer_name (str) β The name of the HuggingFace tokenizer to use to tokenize samples.
max_seq_len (int) β The max sequence length of each token sample.
group_method (str) β How to group text samples into token samples. Currently only supporting
'truncate'
.local (str) β Local filesystem directory where dataset is cached during operation.
remote (str, optional) β Remote directory (S3 or local filesystem) where dataset is stored. Defaults to
None
.split (str, optional) β The dataset split to use, either βtrainβ or βvalβ. Defaults to
None
.shuffle (bool) β Whether to iterate over the samples in randomized order. Defaults to
True
.prefetch (int, optional) β Target number of samples remaining to prefetch while iterating. Defaults to
100_000
.keep_zip (bool, optional) β Whether to keep or delete the compressed file when decompressing downloaded shards. If set to None, keep iff remote is local. Defaults to
None
.retry (int) β Number of download re-attempts before giving up. Defaults to
2
.timeout (float) β Number of seconds to wait for a shard to download before raising an exception. Defaults to
60
.hash (str, optional) β Hash or checksum algorithm to use to validate shards. Defaults to
None
.batch_size (int, optional) β Hint the batch size that will be used on each deviceβs DataLoader. Defaults to
None
.