mikezhu/fineweb-edu-100B-gpt2-tokenized-1024

FineWeb-Edu 100B — GPT-2 tokenized, wrapped to length 1024

Pre-tokenized version of HuggingFaceFW/fineweb-edu (sample-100BT config), prepared for autoregressive language model pretraining.

This is only the training split. Validation is wikitext-103 (not included here).

Tokenization

• Tokenizer: gpt2
• Sequence length: 1024
• Wrap: True — documents are concatenated and chunked into fixed-length sequences
• EOS inserted between documents
• Saved with 🤗 datasets save_to_disk (Arrow shards)

Stats

• Rows: 97,961,640 sequences
• Tokens: ~100.3 B (1024 × rows)
• Shards: 1607 Arrow files under train/ (each ~500 MB)

Schema

{
  "input_ids":      Sequence(int32),    # length 1024
  "attention_mask": Sequence(float32),  # length 1024 (all 1.0 — wrap=True has no padding)
}

Note: attention_mask is stored as float32 because that is what the original training pipeline produced. With wrap=True it is always all-ones and can be regenerated on the fly to save bandwidth/disk.

Loading

from datasets import load_dataset
ds = load_dataset("mikezhu/fineweb-edu-100B-gpt2-tokenized-1024",
                  split="train", streaming=True)

Or download the raw shards and use load_from_disk:

from huggingface_hub import snapshot_download
from datasets import load_from_disk

path = snapshot_download(
    repo_id="mikezhu/fineweb-edu-100B-gpt2-tokenized-1024",
    repo_type="dataset",
    allow_patterns="train/*",
)
ds = load_from_disk(f"{path}/train")

Source & licensing

• Source dataset: HuggingFaceFW/fineweb-edu — ODC-By 1.0. The same license applies to this derivative.
• Tokenizer: gpt2.

Reproduction

Tokenization was done as part of the duo-ar training pipeline (configs/data/fineweb-edu-100B.yaml), equivalent to:

train: fineweb-edu-100B
tokenizer_name_or_path: gpt2
wrap: True
insert_train_eos: True