---
language:
- en
license: odc-by
pretty_name: FineWebEDU Sample Text
size_categories:
- 100K<n<1M
task_categories:
- text-generation
tags:
- text
- web
- educational
- sample
- gzip
---

# Dataset Card for FineWebEDU Sample Text

## Dataset Details

### Dataset Description

FineWebEDU Sample Text is a small derivative dataset intended for coursework and local experimentation. It contains the first 1 GiB of UTF-8 plain text extracted from the beginning of the [`HuggingFaceFW/fineweb-edu`](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu) dataset and distributed here as a single gzip-compressed text file.

- **Curated by:** Zico Kolter (sample packaging); upstream dataset curated by Hugging Face
- **Funded by [optional]:** Not applicable
- **Shared by [optional]:** Zico Kolter
- **Language(s) (NLP):** English
- **License:** Open Data Commons Attribution License (ODC-By) v1.0; use is also subject to Common Crawl Terms of Use

### Dataset Sources

- **Repository:** https://huggingface.co/datasets/zkolter/FineWebEDU-Sample-Text
- **Upstream dataset:** https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu
- **Paper [optional]:** https://arxiv.org/abs/2406.17557

## Uses

### Direct Use

This sample is suitable for small-scale tokenizer experiments, compression and I/O benchmarks, course assignments, and toy language-model preprocessing pipelines where a single local text file is more convenient than streaming parquet data.

### Out-of-Scope Use

This repo is not a faithful substitute for the full FineWeb-Edu release. It should not be used where exact document boundaries, per-record metadata, representative sampling, or complete upstream coverage matter.

## Dataset Structure

- **File:** `finewebedu_first_1gb.txt.gz`
- **Compression:** gzip
- **Uncompressed format:** UTF-8 plain text
- **Uncompressed size:** exactly 1,073,741,824 bytes
- **Compressed size in this repo:** 406,983,843 bytes
- **Approximate number of source documents included:** 221,084
- **Source config/split/shard:** `CC-MAIN-2013-20` / `train` / `data/CC-MAIN-2013-20/train-00000-of-00014.parquet`
- **Splits:** none

This is a raw text corpus dump, not JSONL, CSV, or a line-delimited dataset. Source documents were concatenated in order with a trailing newline after each document, but documents can themselves contain internal newlines. As a result, line boundaries do not reliably correspond to document boundaries.

## Dataset Creation

### Curation Rationale

This sample was created to make a subset of FineWeb-Edu easy to use in local coursework without downloading multi-gigabyte parquet shards or setting up parquet-specific tooling.

### Source Data

#### Data Collection and Processing

The source text came from the public `HuggingFaceFW/fineweb-edu` dataset. The sample was built by reading the `text` column from the first shard of the `CC-MAIN-2013-20` training split and concatenating documents in source order until the output reached exactly 1 GiB.

To satisfy the fixed-size target, the final document was truncated on a UTF-8 character boundary. This means the sample preserves document order from the start of the shard, but the last document is incomplete.

#### Who are the source data producers?

The underlying text originates from public web pages included in Common Crawl and filtered by the FineWeb-Edu curation pipeline. The ultimate source-data producers are the original web authors and publishers.

#### Personal and Sensitive Information

This repo does not add annotations or perform additional privacy filtering beyond the upstream dataset. Because the content is derived from public web pages, it may still contain personal data, copyrighted material references, or other sensitive content present in the original pages.

## Bias, Risks, and Limitations

This sample inherits the biases and risks of Common Crawl and FineWeb-Edu, including uneven topical coverage, language and domain imbalance, and the presence of harmful or low-quality web content despite upstream filtering.

This repo also introduces additional limitations:

- it is a prefix sample from a single shard rather than a representative sample of the full dataset;
- it discards all upstream metadata except the extracted text;
- it stores data in a raw concatenated text format rather than record-oriented examples; and
- the final source document is truncated.

### Recommendations

Use this repo for lightweight experimentation only. For serious model training, evaluation, or dataset analysis, prefer the upstream FineWeb-Edu release and retain per-document metadata and provenance.

## Citation

If you use this sample, cite the upstream FineWeb-Edu dataset and paper:

**BibTeX:**

```bibtex
@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}
}
```

## More Information

This repo is a convenience sample derived from the upstream dataset for classroom use. It is not an official Hugging Face release.

## Dataset Card Authors

Zico Kolter

## Dataset Card Contact

zkolter@cs.cmu.edu