TabLib

Name: tablib-v1-full
Creator: Approximate Labs
License: https://choosealicense.com/licenses/other/

A minimally-preprocessed dataset of 627M tables (69 TiB) extracted from HTML, PDF, CSV, TSV, Excel, and SQLite files from GitHub and Common Crawl.

This includes 867B tokens of "context metadata": each table includes provenance information and table context such as filename, text before/after, HTML metadata, etc.

A smaller 0.1% sample of this dataset can be found here.

For more information, read the paper & announcement blog.

Dataset Details

Sources

GitHub: nearly all public GitHub repositories
Common Crawl: the CC-MAIN-2023-23 crawl

Reading Tables

Tables are stored as serialized Arrow bytes in the arrow_bytes column. To read these, you will need to deserialize the bytes:

import datasets
import pyarrow as pa

# load a single file of the dataset
ds = datasets.load_dataset(
    'approximatelabs/tablib-v1-full',
    data_files='tablib/job=github_000005/batch=000001/part=000001/manifest.parquet',
    token='...',
)

df = ds['train'].to_pandas()

tables = [pa.RecordBatchStreamReader(b).read_all() for b in df['arrow_bytes']]

Licensing

This dataset is intended for research use only.

For specific licensing information, refer to the license of the specific datum being used.

Contact

If you have any questions, comments, or concerns about licensing, pii, etc. please contact using this form.

Approximate Labs

TabLib is a project from Approximate Labs. Find us on Twitter, Github, Linkedin, and Discord.

Citations

If you use TabLib for any of your research, please cite the TabLib paper:

@misc{eggert2023tablib,
      title={TabLib: A Dataset of 627M Tables with Context}, 
      author={Gus Eggert and Kevin Huo and Mike Biven and Justin Waugh},
      year={2023},
      eprint={2310.07875},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Datasets:

approximatelabs
/

tablib-v1-full

You need to agree to share your contact information to access this dataset