README.md

metadata

license: mit

Stage 1 Packed Pretraining Dataset

This dataset contains preprocessed and token-packed .bin files intended for use in pretraining a decoder-only Transformer language model.

Dataset Contents

• Each .bin file contains a fixed number of samples, where each sample is exactly 8192 tokens long.
• Samples are grouped into batches of 125 samples, totaling 1.024 million tokens per batch.
• Each file (called a "block") contains 62500 samples (approximately 512 million tokens).
• All samples are tokenized using the GPT2TokenizerFast from Hugging Face Transformers.

Structure

• Format: Binary files (int32) containing token IDs.
• File naming: stage1_block_0000.bin, stage1_block_0001.bin, etc.
• Tokenizer: GPT2TokenizerFast with eos_token used as a separator and padding token.
• Context length: 8192 tokens per sample.

Source Datasets

Tokens were drawn from a diverse mix of high-quality open datasets:

• C4 (en)
• Wikipedia (2023/11 dump)
• OpenWebText
• CCNews
• Gutenberg
• arXiv
• BookCorpus Open
• S2ORC
• TriviaQA
• PAQ
• Natural Questions

Each dataset was assigned a token quota to ensure a balanced mix.

Preprocessing & Packing Strategy

• Samples were streamed using Hugging Face Datasets with shuffling.
• Texts were tokenized, filtered using a garbage filter, and concatenated with separator tokens.
• Samples were packed into fixed-length chunks of 8192 tokens.
• Leftover tokens from one batch are carried forward to the next to ensure no token duplication or loss.

Garbage Filtering Heuristics:

• Removed texts with:
  • Too few words or characters.
  • High symbol-to-alphanumeric ratio.
  • Excessive character repetition.
  • Very low word diversity.

Usage Example

You can load and decode tokens using PyTorch:

import torch
from transformers import GPT2TokenizerFast

tokenizer = GPT2TokenizerFast.from_pretrained("gpt2")
with open("stage1_block_0000.bin", "rb") as f:
    tokens = torch.frombuffer(f.read(), dtype=torch.int32)
    sample = tokens[:8192].tolist()
    text = tokenizer.decode(sample)
    print(text)