Datasets:

CMSManhattan
/

wiki13_tokenized_precision

license: cc-by-sa-4.0
language:
  - en
  - zh
  - fr
  - es
  - pt
  - de
  - it
  - ru
  - ja
  - ko
  - vi
  - th
  - ar
tags:
  - wikipedia
  - pretraining
  - tokenized

wiki13_tokenized_precision

Wikipedia in 13 languages, tokenized with CMSManhattan/JiRackPrecisionTokenizer and packed into 2048-token blocks (no BOS, EOS between articles). Token shares: en 50%, the other 12 languages 4.17% each. 97 train shards (wiki_train_shard_NNNN.pt, 5000 blocks each, ~0.99B tokens) + wiki_val.pt (512 blocks). Format: torch.load(f)["input_ids"] = list of LongTensor(2048).

Source: Wikipedia (https://www.wikipedia.org), licensed CC BY-SA 4.0. This dataset is a derivative (tokenized) and is distributed under the same licence.