krisbailey/fineweb-edu-1B

FineWeb-Edu 1B

Dataset Description

FineWeb-Edu 1B is a high-quality, stratified subset of the HuggingFaceFW/fineweb-edu dataset. It contains approximately 1 billion tokens of educational web text, carefully sampled to preserve the original distribution of source data (CommonCrawl dumps).

This dataset provides an accessible, lightweight alternative to the larger FineWeb-Edu subsets (like sample-10BT or sample-100BT) while maintaining the same data diversity and quality characteristics. It is ideal for rapid prototyping, educational purposes, and testing models on a smaller scale without downloading terabytes of data.

Dataset Details

• Total Tokens: ~1.001 Billion
• Source: HuggingFaceFW/fineweb-edu (split: sample-10BT)
• Language: English
• Format: Apache Parquet
• Producer: Kris Bailey (kris@krisbailey.com)

Motivation

The original FineWeb-Edu dataset is excellent but massive. I created this version because I found there wasn't a good intermediate dataset available between the 10B token version and the very small 1M token version. Even the sample-10BT subset can be unwieldy for quick experiments.

I created this 1B token version to offer a "mini" variant that:

1. Is Easy to Access: Small enough to download and load into memory quickly.
2. Preserves Diversity: Maintains the exact same proportion of data from different distinct CommonCrawl dumps as the original dataset, ensuring the temporal and structural variety is not lost during downsampling.
3. Maintains Quality: Derived directly from the high-quality educational filtered data of FineWeb-Edu.

Dataset Creation Process

The detailed steps taken to create this dataset are as follows:

1. Source Selection

We started with the sample-10BT subset of HuggingFaceFW/fineweb-edu. This subset itself is a random sample of the larger dataset.

2. Token Calculation

We scanned the source dataset to calculate the exact total number of tokens:

• Source Total Tokens: ~9.97 Billion

3. Probabilistic Sampling

To reduce the size to 1 Billion tokens while strictly preserving the distribution of the dump column (the identifier for different CommonCrawl crawls), we employed a probabilistic sampling method rather than a simple top-N truncation.

We calculated a sampling probability $P$: $$P=\frac{\text{Target Tokens}}{\text{Total Tokens}}\approx \frac{1,000,000,000}{9,972,918,177}\approx 0.1002P\; =\; \backslash frac\{\backslash text\{Target\; Tokens\}\}\{\backslash text\{Total\; Tokens\}\}\; \backslash approx\; \backslash frac\{1,000,000,000\}\{9,972,918,177\}\; \backslash approx\; 0.1002$$

Records were selected where random.random() < P. This method ensures that every sub-group (dump) is sampled at the same rate, statistically preserving the original percentage breakdown of the data.

4. Verification

After sampling, we verified the integrity of the new subset:

• Target Size: 1,000,000,000 tokens
• Actual Size: 1,001,638,506 tokens (+0.16%)
• Distribution Check: Validated that the relative percentage of each CommonCrawl dump in this 1B subset matches the original sample-10BT distribution.

5. Format Conversion

The dataset was converted from Apache Arrow to Apache Parquet to optimize storage and transfer efficiency.

• Original Size (Arrow): ~4.9 GB
• New Size (Parquet): ~2.8 GB
• Compression: ~43% reduction in storage footprint.
• Verification: Row counts were strictly verified against the original Arrow files to ensure zero data loss.

fast-usage

from datasets import load_dataset

# Load the dataset
ds = load_dataset("krisbailey/fineweb-edu-1B", split="train")

# Print an example
print(ds[0])

Citation

If you use this dataset, please cite the original FineWeb-Edu paper:

@article{lmazmumi2024finewebedu,
  title={FineWeb-Edu},
  author={Lmazmumi,  Wissam and  Genedy,  Ahmed and  Cochran,  Sean and  Taylor,  Ross and  Bistline,  Graeme and  Sunny,  Namrat and  Olson,  Carly and  Sadeq,  Sultan and  Vaidya,  Nilesh and  Sarti,  Paul  and others},
  journal={Hugging Face},
  year={2024},
  url={https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu}
}

License

This dataset is released under the Open Data Commons Attribution License (ODC-By) v1.0. Usage is also subject to CommonCrawl's Terms of Use.