Hugging Face's logo

Datasets:
cis-lmu
/
glotlid-corpus

Modalities:
Text
Formats:
text
Size:
100M - 1B
ArXiv:
Libraries:
Datasets
License:
Dataset card Viewer Files Community
27

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

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

By filling the form below, you understand that only the metadata and the annotations of GlotLID have a cc0-1.0 license, and that the rest of the content have each different licenses for which the authors of GlotLID do not hold any copyright whatsoever. Instead of redistributing the data, you can always link to it here or to the GlotLID paper. We maintain the data by preserving previous versions, adding new languages in newer versions, and correcting any mistakes in the metadata. Noted that some parts of the data from certain sources cannot be redistributed publicly and are provided here solely for research purposes. Requests are manually reviewed, and access grants take 1-3 days. We are not obliged to provide access to all who fill the form. The sources of all data are open (see readme), and anyone can obtain them with some effort.

Log in or Sign Up to review the conditions and access this dataset content.

GlotLID Corpus

This is the corpus used to train the open GlotLID model, a language identification model capable of classifying ~2000 labels. The list of raw texts supported in this repository is openly available here: https://github.com/cisnlp/GlotLID/blob/main/sources.md

Download

The recommended way to download the data is through the Huggingface Hub:

pip install huggingface_hub

For faster downloads, make sure to install the additional transfer package and set the environment variable:

pip install huggingface_hub[hf_transfer]
export HF_HUB_ENABLE_HF_TRANSFER=1

After granting access to this repository, you can use your HF token to download the data: Huggingface Tokens Replace your token in with "hf_xxXxXxXXXXxxxxxXXxxxxxXXXXXXXxXXxx"

To download specific language data (bal_Arab is Balochi language written in Arabic script):

from huggingface_hub import snapshot_download

folder = snapshot_download(
    "cis-lmu/glotlid-corpus", 
    repo_type="dataset",
    local_dir="./path/to/glotlid-corpus/",
    # Replace "v3.1/bal_Arab/*" with the path for any other language available in the dataset
    allow_patterns="v3.1/bal_Arab/*",
    token="hf_xxXxXxXXXXxxxxxXXxxxxxXXXXXXXxXXxx"
)

To download the whole dataset (v3.1):

from huggingface_hub import snapshot_download

folder = snapshot_download(
    "cis-lmu/glotlid-corpus", 
    repo_type="dataset",
    local_dir="./path/to/glotlid-corpus/",
    allow_patterns="v3.1/*",
    token="hf_xxXxXxXXXXxxxxxXXxxxxxXXXXXXXxXXxx"
)

Data Format

GlotLID v3.1 is the collection of data used to train GlotLID v3, with some edits (v3.1). Each language_script has a folder; for example, v3.1/bal_Arab is for Balochi written in the Arabic script. In each language folder, you will find .txt files. For example, in v3.1/bal_Arab, you will find bal_Arab_GlotSparse.txt, indicating that this data comes from the GlotSparse source. Another example is bal_Arab_bloombooks+bcc.txt, indicating that this data comes from the Bloombooks source, but in Bloombooks, it was introduced with bcc language label as metadata. Here, we change it to its macro language code, bal, for better classification.

The codes we use for representing the languages are 3-letter ISO 639 codes: bal: https://iso639-3.sil.org/code/bal For the scripts, we use ISO 15924: https://de.wikipedia.org/wiki/ISO_15924 For example, Arab represents Arabic.

License

Only the metadata and annotations of GlotLID are licensed under CC0-1.0. The rest of the content is under various licenses, and the authors of GlotLID do not hold any copyright for it. Instead of redistributing the data, you can link to it here or to the GlotLID paper.

If you use the data from this repository, you consent to this form:

By filling the form below, you understand that only the metadata and the
annotations of GlotLID have a cc0-1.0 license, and that the rest of the
content have each different licenses for which the authors of GlotLID **do not** hold any copyright
whatsoever. Instead of redistributing the data, you can always link to it here or to the GlotLID paper.
We maintain the data by preserving previous versions, adding new languages
in newer versions, and correcting any mistakes in the metadata. Noted that some parts of the data from
certain sources cannot be redistributed publicly and are provided here solely for research purposes.
Requests are manually reviewed, and access grants take **1-3 days**. We are not obliged to provide access to all who fill the form. The sources of all data are open (see readme), and anyone can obtain them with some effort.
Name: text
Email (institution/company): text 
Affiliation: text
Country: text
Usecase (be specific as much as possible): text
I do not redistribute this data unless I am certain that the part I am redistributing can be legally redistributed, I respect the license of each source and am responsible for verifying the license, I use the data only for the described use case: checkbox

Citation

If you find the GlotLID model, list of data sources or our collection of texts here useful, please cite the following paper:

@inproceedings{
  kargaran2023glotlid,
  title={{GlotLID}: Language Identification for Low-Resource Languages},
  author={Kargaran, Amir Hossein and Imani, Ayyoob and Yvon, Fran{\c{c}}ois and Sch{\"u}tze, Hinrich},
  booktitle={The 2023 Conference on Empirical Methods in Natural Language Processing},
  year={2023},
  url={https://openreview.net/forum?id=dl4e3EBz5j}
}
Downloads last month
3,627

Models trained or fine-tuned on cis-lmu/glotlid-corpus