meryyllebr543/ultrafineweb-100B-tokens

Tokenized Ultra-FineWeb (100B English Tokens)

This repository provides a tokenized version of the English split of the openbmb/Ultra-FineWeb dataset, prepared for large-scale language model training. The dataset consists of 100 billion high-quality tokens, processed with a custom tokenizer.

The data is sharded into 100 files, each containing exactly 1 billion tokens, making it easy to stream and use in distributed training setups.

---

Dataset Details

• Source Dataset: openbmb/Ultra-FineWeb
• Language: English
• Total Tokens: 100,000,000,000
• Data Format: Sharded NumPy arrays
• Shard Count: 100 files (shard_0000.npy to shard_0099.npy)
• Tokens per Shard: 1,000,000,000
• Data Type: numpy.uint32
• Tokenizer: Custom BPE tokenizer (see tokenizer.json for details)
• Vocabulary Size: 65,536

---

Usage

You can easily load any shard using NumPy or stream the entire dataset using the datasets library.

Loading a Single Shard

To load a specific shard, use the following Python code:

import numpy as np
from huggingface_hub import hf_hub_download

# Download and load a specific shard
file_path = hf_hub_download(
    repo_id="meryyllebr543/ultrafineweb-100B-tokens", 
    filename="shards/shard_0000.npy"
)
tokens = np.load(file_path)

print(f"Loaded shard with {len(tokens):,} tokens.")
# Expected output: Loaded shard with 1,000,000,000 tokens.

Streaming with the datasets library

This repository is structured to be compatible with the datasets library for streaming.

from datasets import load_dataset

# Stream the dataset (this is memory-efficient)
dataset = load_dataset("meryyllebr543/ultrafineweb-100B-tokens", streaming=True, split="train")

for item in dataset:
    # Each 'item' will be a dictionary containing a batch of tokens
    # The structure will depend on how the data is configured in your repo
    print(item)
    break 

---

About the Original Ultra-FineWeb Dataset

This tokenized dataset is derived from Ultra-FineWeb, which is a large-scale, high-quality, and efficiently-filtered dataset. It was created by applying an efficient, verification-based filtering pipeline to the FineWeb dataset. Ultra-FineWeb serves as a core pre-training web dataset for the MiniCPM Series models.

For a complete understanding of the data filtering, verification, and evaluation, please refer to the official Ultra-FineWeb technical report.

---

License

The scripts used to generate this dataset are released under the Apache 2.0 license. The dataset itself is a derivative of openbmb/Ultra-FineWeb, which is also licensed under Apache 2.0. Following the original authors' guidelines, users of this dataset should also be aware of the licenses of the underlying data sources used to create FineWeb.

---

Citation

If you use this dataset in your work, please be sure to cite the original authors of the Ultra-FineWeb dataset:

@misc{wang2025ultrafineweb,
  title={{Ultra-FineWeb}: Efficient Data Filtering and Verification for High-Quality LLM Training Data},
  author={Yudong Wang and Zixuan Fu and Jie Cai and Peijun Tang and Hongya Lyu and Yewei Fang and Zhi Zheng and Jie Zhou and Guoyang Zeng and Chaojun Xiao and Xu Han and Zhiyuan Liu},
  year={2025},
  eprint={2505.05427},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
}