beshiribrahim's picture
Update README.md
2e23b28 verified
metadata
license: cc-by-sa-4.0
pretty_name: Tigre language
language:
  - tig

Tigre Low-Resource Language Resource Collection

Overview

This repository introduces the Monolingual Text component of the Tigre language resource collection. Tigre is an under-resourced South Semitic language within the Afro-Asiatic family. This dataset provides a large, clean text corpus essential for training foundational models such as Language Models (LMs) and word embeddings. The goal of Tigre-Data 1.0 is to accelerate research in low-resource NLP and morphologically rich language modeling.


Included Data & Statistics

Data Modalities

This repository contains only the Monolingual Text data modality.

Dataset Statistics

The corpus was tokenized using a simple whitespace tokenizer to determine the core metrics below.

Statistic Value
Total Number of Examples (Rows) 490,032
Total Number of Tokens 14,700,960
Vocabulary Size (Unique Tokens) 760,384
Average Example Length 30.00 tokens/row

Dataset Structure

The dataset is provided in the Parquet format, which is easily streamed and loaded using the Hugging Face datasets library.

tigre-data-monolingual-text/
├── README.md
├── data.parquet
└── arrow_format/
    └── train/
        ├── data-00000-of-00001.arrow
        ├── dataset_info.json
        └── state.json

Data Provenance & Methodology

Sources

The monolingual text corpus was compiled from diverse sources to maximize coverage:

  • Books
  • News articles
  • Web content
  • Wikipedia

Data Curation & Preprocessing

  • Preprocessing: The data underwent a light cleanup of data to remove non text binaries.
  • Orthographic Normalization: The original corpus was normalized to ensure consistent Ge'ez script usage.
  • Text Cleaning: Steps such as deduplication and boilerplate removal were applied to improve corpus quality (details available in the associated data paper).

Bias, Risks & Known Limitations

The data collection process was designed to be broad; however, inherited biases from the original sources are present:

  • Domain Bias: The sources (news articles, history books, poems, culture-related texts) mean the corpus may overrepresent formal and historical language and underrepresent informal or conversational Tigre.
  • Linguistic Bias: Any inherent orthographic variation or dialectal representation present in the original source materials is inherited by this dataset.

How to Download & Load the Dataset

The dataset can be easily loaded using the Hugging Face Hub client library:

from datasets import load_dataset

dataset_name = "BeitTigreAI/tigre-data-monolingual-text"

# Load the full dataset (the default split is 'train')
ds = load_dataset(dataset_name, split="train")

# Example: Display the number of rows and the first example
print(f"Total rows loaded: {len(ds)}")
print(ds[0])

```python

## Licensing

CC-BY-SA-4.0

## Citation

If you use this resource in your work, please cite the repository by referencing its Hugging Face entry:

### Recommended Citation Format:

- Repository Name: Tigre Monolingual Text Dataset
- Organization: BeitTigreAI
- URL: https://huggingface.co/datasets/BeitTigreAI/tigre-data-monolingual-text