anisoleai/fineweb-tokenized

FineWeb Tokenized Corpus (AnisoleAI)

Overview

This repository provides a large-scale pre-tokenized version of the FineWeb dataset designed for efficient training of language models.

The dataset contains text from the FineWeb corpus that has been tokenized using a SentencePiece tokenizer.
Tokens are stored in a compact uint16 format for efficient storage and high-throughput training.

Each dataset record contains a flat array of token IDs representing a continuous tokenized text sequence.

This format allows:

• fast loading
• minimal memory overhead
• efficient distributed training
• direct compatibility with LLM training pipelines

No semantic modifications were made to the original FineWeb dataset.
The text was only tokenized and serialized into shard files.

---

Dataset Structure

The dataset is organized into multiple shard directories:

data_1/shard-00000.parquet
data_1/shard-00001.parquet
...

data_2/shard-00000.parquet
...

...

data_20/shard-XXXXX.parquet

Each shard contains:

token_ids: uint16[]

Each record stores a contiguous tokenized segment that can be used directly for model training.

---

Loading the Dataset

You can load the dataset using the HuggingFace datasets library.

from datasets import load_dataset

dataset = load_dataset(
    "anisoleai/fineweb-tokenized",
    split="shard"
)

sample = dataset[0]["token_ids"]

print("Number of tokens:", len(sample))

The dataset supports:

• streaming
• distributed loading
• partial downloads

---

Loading the Tokenizer

The tokenizer used to generate the corpus is included in this repository.

import sentencepiece as spm
from huggingface_hub import hf_hub_download

model_path = hf_hub_download(
    repo_id="anisoleai/fineweb-tokenized",
    filename="tokenizer.model",
    repo_type="dataset"
)

sp = spm.SentencePieceProcessor(model_file=model_path)

print("Vocabulary size:", sp.get_piece_size())

print(sp.decode([1, 10, 20, 30]))

---

Intended Use

This dataset is intended for:

• large language model pretraining
• tokenizer benchmarking
• distributed LLM training pipelines
• academic AI research
• commercial AI development

The shard-based structure allows scalable multi-worker training pipelines.

---

Source Dataset

Original dataset:

FineWeb
https://huggingface.co/datasets/HuggingFaceFW/fineweb

FineWeb is a large-scale filtered web corpus designed for training language models.

---

License

This dataset follows the license of the original dataset:

Open Data Commons Attribution License (ODC-BY) v1.0

https://opendatacommons.org/licenses/by/1-0/

---

Attribution

If you use this dataset, please attribute:

• the creators of the FineWeb dataset
• AnisoleAI for the tokenization pipeline and dataset preparation

---

Notes

• The dataset contains token IDs only.
• Original raw text is not included.
• Token IDs correspond to the included SentencePiece tokenizer.