gpjt/openwebtext-gpt2-tokens-DEPRECATED

This dataset has been deprecated due to contamination with documents that exist in my test/validation dataset

OpenWebText GPT-2 Tokens

This dataset is 10B GPT-2 tokens, generated by oversampling OpenWebText, stored as a single uint16 tensor.

The source dataset was sampled randomly, and the text for each sampled record was tokenized using tiktoken, then they were all concatenated together, with <|endoftext|> tokens separating them.

As there are somewhat fewer than 10B tokens in the source dataset, it was oversampled by a small amount.

Code:

• The download/sample/upload code
• The config for this dataset

Based on

This dataset is derived from Aaron Gokaslan's OpenWebText dataset:

• Source: OpenWebText (by Aaron Gokaslan)
• Processing:
  • Input text sampled from OpenWebText
  • Tokenized using tiktoken with GPT-2 encoding
  • <|endoftext|> inserted between documents
  • Stored as uint16 safetensors

File format

• train.safetensors – token stream for training

The file contains a single tensor:

{"tokens": torch.uint16[...]}