BeitTigreAI/tigre-data-parallel-multilingual

Tigre Parallel Multilingual Dataset (Tigre-Data 1.0)

Overview

This repository introduces the Parallel Multilingual Text component of the Tigre language resource collection. Tigre is an under-resourced South Semitic language within the Afro-Asiatic family.

The goal of Tigre-Data 1.0 is to accelerate research in low-resource NLP and morphologically rich language modeling. This dataset provides a clean, high-quality parallel corpus essential for developing and evaluating Machine Translation (MT) systems for Tigre.

Data Source & Licensing

The parallel sentences in this dataset originate from Tatoeba.org, a community-driven multilingual corpus released under the CC-BY 2.0 license. All Tigre translations were contributed by native-speaking members of the Tigre diaspora, reflecting years of collective volunteer effort to expand the language’s digital presence. This contribution is significant, as, by the end of 2025, the Tigre language on Tatoeba.org is supported by 77 registered translators working across more than ten other languages, resulting in a larger sentence pool than 93% of the 429 hosted languages.

Included Data & Structure

Data Modalities

This repository contains the Parallel Multilingual Text modality.

Dataset Structure

The dataset is provided in Parquet format, compatible with the Hugging Face datasets library.

tigre-data-parallel-multilingual/
├── README.md
└── tigre-data-parallel-multilingual.parquet

Dataset Statistics and Metrics

The dataset contains 329,554 parallel sentences across seven target languages.

Corpus Totals

Statistic	Value
Total Sentences in Corpus	329,554
Total Source Words (Tigre)	1,072,951
Total Target Words	1,240,407
Avg Sentence Length (Tigre)	3.26 words
Avg Sentence Length (Target)	3.76 words

Language Pair Distribution (Tigre → Target Language)

Target Language	Total Sentences	Percentage	Tigre Words	Avg Src Words/Sentence	Target Words	Avg Tgt Words/Sentence
ara_Arab (Arabic)	49,964	15.16%	172,645	3.46	160,675	3.22
deu_Latn (German)	80,120	24.31%	258,160	3.22	321,966	4.02
eng_Latn (English)	156,206	47.40%	508,318	3.25	610,321	3.91
nno_Latn (Norwegian Nynorsk)	6,001	1.82%	14,055	2.34	14,619	2.44
nob_Latn (Norwegian Bokmål)	8,222	2.49%	21,567	2.62	23,481	2.86
swe_Latn (Swedish)	28,554	8.66%	96,439	3.38	107,562	3.77
tir_Ethi (Tigrinya)	487	0.15%	1,767	3.63	1,783	3.66

How to Download & Load the Dataset

from datasets import load_dataset
import pandas as pd

dataset = load_dataset(
    "BeitTigreAI/tigre-data-parallel-multilingual",
    data_files={
        "train": "train.parquet",
        "validation": "validation.parquet"
    }
)

print("Dataset Info:")
print(dataset)

df_train = dataset["train"].to_pandas()
lang_counts = df_train["tgt_lang"].value_counts().to_frame().reset_index()
lang_counts.columns = ["tgt_lang", "count"]
lang_counts["percentage"] = (lang_counts["count"] / lang_counts["count"].sum() * 100).round(2)
print(lang_counts.to_string(index=False))

Licensing

This dataset is released under the CC-BY-SA-4.0 license.

Citation

If you use this resource, please cite the dataset using its Hugging Face entry.

Repository: Tigre Parallel Multilingual Dataset Organization: BeitTigreAI URL: https://huggingface.co/datasets/BeitTigreAI/tigre-data-parallel-multilingual