gpjt/fineweb-gpt2-tokens

FineWeb GPT-2 Tokens

This is all of the 10B sample of FineWeb, split into 99% training data, 1% validation data, each converted to a single uint16 tensor of GPT-2 tokens. The text for each record in the original dataset was tokenized using tiktoken, then they were all concatenated together, with <|endoftext|> tokens separating them.

Code:

• Script to download the original dataset
• Script to do the tokenization

Based on

This dataset is derived from the original FineWeb dataset:

• Source: FineWeb (by Hugging Face)
• Processing:
  • Input text from FineWeb 10B sample
  • Tokenized using tiktoken with GPT-2 encoding
  • <|endoftext|> inserted between documents
  • Stored as uint16 safetensors

File format

• train.safetensors – token stream for training
• validation.safetensors – token stream for validation

Each file contains a single tensor:

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