README.md

---
language:
- en
license: odc-by
pretty_name: TaoData
size_categories:
- 1M<n<10M
task_categories:
- language-modeling
task_ids:
- language-modeling
tags:
- pretraining
- web
- fineweb-edu
- taodata
configs:
- config_name: default
  data_files:
  - split: train
    path: pretrain.jsonl
---

# TaoData

## Dataset Summary

TaoData is a large-scale English pretraining dataset prepared for the Taotern model family. It is a processed subset of [HuggingFaceFW/fineweb-edu](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu), filtered and formatted with the [TaoData framework](https://github.com/Taotern/TaoData) using the [`taodata.yaml`](https://github.com/Taotern/TaoData/blob/main/taodata/config/taodata.yaml) configuration.

This release contains about 6.7 million samples in JSONL format and is intended as one of the main pretraining corpora used by Taotern models.

## Supported Tasks

- Causal language modeling
- Continued pretraining
- Data mixture construction for foundation models

## Languages

- English

## Dataset Structure

### Data Instances

Each row is a JSON object. Example:

```json
{
  "text": "Average life span in the wild: 12 years...",
  "id": "<urn:uuid:9c71b6db-6728-48b5-96b5-05fbc0b5bb4f>",
  "dump": "CC-MAIN-2013-20",
  "url": "http://amazingpicturesoftheanimals.blogspot.com/2012/05/mellers-chameleon-facts-pictures.html",
  "date": null,
  "file_path": "s3://commoncrawl/...",
  "language": "en",
  "language_score": 0.9396425485610962,
  "token_count": 359,
  "score": 3.40625,
  "int_score": 3
}
```

### Data Fields

- `text`: Cleaned document text used for pretraining.
- `id`: Sample identifier inherited from the upstream dataset.
- `dump`: Common Crawl dump identifier.
- `url`: Source URL from the crawl.
- `date`: Source date when available.
- `file_path`: Original Common Crawl WARC path.
- `language`: Language tag.
- `language_score`: Upstream language confidence score.
- `token_count`: Token count metadata.
- `score`: Quality score metadata.
- `int_score`: Integer-binned quality score.

### Data Splits

- `train`: ~6.7M samples

## Source Data

This dataset is derived from:

- [HuggingFaceFW/fineweb-edu](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu)

The upstream dataset is a large English web corpus derived from Common Crawl and released by Hugging Face. TaoData applies the TaoData preprocessing pipeline to produce a cleaned training subset suitable for large-scale language model pretraining.

## Processing

The dataset was processed with the TaoData framework using:

- [`taodata/config/taodata.yaml`](https://github.com/Taotern/TaoData/blob/main/taodata/config/taodata.yaml)

At a high level, this release is intended to be a clean, training-ready subset for Taotern model development.

## Intended Use

TaoData is designed for:

- Pretraining or continued pretraining of large language models
- Training data ablations and mixture experiments
- Research on web-scale text filtering and corpus construction

## Limitations

- This is web-derived text and may still contain factual errors, bias, noise, or unwanted content.
- The dataset inherits limitations, source coverage, and potential artifacts from FineWeb-Edu and Common Crawl.
- It is intended for research and model training, not as a source of verified facts.

## Licensing

This dataset is derived from FineWeb-Edu, which is published on Hugging Face under the `odc-by` license at the time of writing. Users should review the upstream dataset card and ensure their use complies with all applicable upstream terms and attribution requirements.

## Citation

If you use this dataset, please cite the upstream dataset and the TaoData processing framework as appropriate:

```bibtex
@misc{fineweb_edu,
  title = {FineWeb-Edu},
  howpublished = {\url{https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu}}
}

@misc{taodata,
  title = {TaoData},
  howpublished = {\url{https://github.com/Taotern/TaoData}}
}
```