|
|
--- |
|
|
language: |
|
|
- yor |
|
|
- hau |
|
|
- ibo |
|
|
- eng |
|
|
license: mit |
|
|
task_ids: |
|
|
- translation |
|
|
pretty_name: LangBridge Multilingual Translation Dataset |
|
|
size_categories: |
|
|
- 1M<n<10M |
|
|
source_datasets: |
|
|
- original |
|
|
tags: |
|
|
- nigerian-languages |
|
|
- multilingual |
|
|
- translation |
|
|
--- |
|
|
# LangBridge Multilingual Translation Dataset 🇳🇬 |
|
|
This dataset contains parallel translation pairs between **Yoruba (yor_Latn)**, **Hausa (hau_Latn)**, **Igbo (ibo_Latn)**, and **English (eng_Latn)**. |
|
|
It was designed for fine-tuning multilingual translation models such as [facebook/nllb-200-distilled-600M](https://huggingface.co/facebook/nllb-200-distilled-600M), in order for it to achieve state of the art performance at translation tasks. |
|
|
--- |
|
|
|
|
|
## 📦 Dataset Summary |
|
|
The **LangBridge dataset** provides approximately **3.5 million** translation pairs across three Nigerian languages and English. |
|
|
|
|
|
## Training data |
|
|
1038492 ---> Hausa |
|
|
1239879 ---> Igbo |
|
|
1271811 ---> Yoruba |
|
|
|
|
|
## Testing and validation data |
|
|
Please note that, the validation and testing data were made this way because of the proportion of line of each language. This was done to achieve equilibrium. |
|
|
58721 ---> Hausa |
|
|
93000 ---> Igbo |
|
|
100000 ---> Yoruba |
|
|
|
|
|
It aims to improve translation quality for underrepresented African languages in NMT (Neural Machine Translation) systems. |
|
|
|
|
|
Each record in the dataset follows the Ideal translation JSON structure: |
|
|
|
|
|
```json |
|
|
{ |
|
|
"translation": { |
|
|
"yor_Latn": "Mo feran ede Yoruba.", |
|
|
"eng_Latn": "I love the Yoruba language." |
|
|
} |
|
|
} |
|
|
{"translation": |
|
|
{ |
|
|
"ibo_Latn": "Nd\u1ecb l\u1ecdghachiri n\u2019enwegh\u1ecb mgbagha g\u1ee5nyere Husseina Bangshika d\u1ecb ka osote onye isi ala, Steve Nwosu d\u1ecb ka Treasurer, Kabir Alabi Garba (osote onye isi ala West), Gabriel Akinadewo Folajimi (onye enyemaka odeakw\u1ee5kw\u1ecd), Umoru Ibrahim (Vice President North), Muhammad Sanusi na Paulyn Owhievbie Ugbodaga (ha ab\u1ee5\u1ecd Standing Committee). North). Ach\u1ee5p\u1ee5r\u1ee5 Messrs Akinreti Qasim (\u1ecckwa odeakw\u1ee5kw\u1ecd ukwu) na Ademola Adegbamigbe (K\u1ecdmitii guzoro \u1ecdt\u1ecd) maka akw\u1ee5kw\u1ecd na-ezighi ezi..", |
|
|
"eng_Latn": "Those who do not have a quarrel with Husseina Bangshika, like the current President, Steve Nwosu as Treasurer, Kabir Alabi Garba (current President West), Gabriel Akinadewo Folajimi (assistant secretary), Umoru Ibrahim (Vice President North), Muhammad Sanusi and Paulyn Owhievbie Ugbodaga (both Standing Committee members). Also, Messrs Akinreti Qasim (Assistant Secretary General) and Ad" |
|
|
|
|
|
}} |
|
|
As you can see the utf-encoding of some of the rather complex tokens like A, E, I, Ị, O, Ọ, U, Ụ have already been transformed into UTF-8 encoding and this is perfectly fine. Your compile as well as your tranformer would not face any troubles during traning. |
|
|
``` |
