VikramLex/fineweb-edu-llama3-tokenized

FineWeb-Edu, tokenized with Llama 3.1

FineWeb-Edu as ready-to-train rows of 4,096 tokens.

Set	Tokens	Rows	Size
100B	99,999,997,952	24,414,062	400 GB, in 25 parts
3B	2,999,996,416	732,421	12 GB

The 3B set is the first 732,421 rows of the 100B set.

Format

• Tokens are little-endian int32 from the Llama 3.1 tokenizer (unsloth/Meta-Llama-3.1-8B, revision e9a141a2).
• Each document is followed by <|end_of_text|> (id 128001), with no beginning-of-text token. Documents are joined and cut into rows of 4,096 tokens, so a document can continue into the next row. There is no padding.
• Source: FineWeb-Edu revision 87f09149, all of sample/100BT, then the first five files and part of the sixth from CC-MAIN-2025-26.
• manifest.json lists the source files and the sha256 of every file here. The .idx files hold row offsets; every row is 4,096 tokens, so you can ignore them.

Use

hf download VikramLex/fineweb-edu-llama3-tokenized --repo-type dataset --include "3B/*" --local-dir fineweb-edu

import numpy as np

rows = np.memmap("fineweb-edu/3B/fineweb_edu_3B.bin", dtype="<i4", mode="r").reshape(-1, 4096)

Each 100B part holds 1,000,000 whole rows (the last holds 414,062), so a part works on its own. To rebuild the single 100B file, run cat fineweb_edu_100B.bin.part-* > fineweb_edu_100B.bin; its sha256 is 85a1e8f5ba2a8804ee7c7865ed093bda424e24f62e174e7d6b10d0d4a08417ed.

License

Derived from FineWeb-Edu, which is released under ODC-By v1.0 and subject to Common Crawl's Terms of Use. The Llama 3.1 tokenizer is not included and has its own license.

Citation

@misc{lozhkov2024fineweb-edu,
    author       = { Lozhkov, Anton and Ben Allal, Loubna and von Werra, Leandro and Wolf, Thomas },
    title        = { FineWeb-Edu: the Finest Collection of Educational Content },
    year         = 2024,
    url          = { https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu },
    doi          = { 10.57967/hf/2497 },
    publisher    = { Hugging Face }
}