ghananlpcommunity/pristine-twi-english-parallel-sentences

Pristine Twi–English Parallel Sentences Dataset

A massive, sentence-level, deduplicated Twi ↔ English parallel dataset optimized specifically for training State-of-the-Art (SOTA) Machine Translation models.

This dataset is derived from the Pristine Twi-English dataset by the Ghana NLP Community. While the original dataset consists of full passages spanning various styles (Monologue, Narrative, Dialogue, Storyful), this version has been meticulously processed, split into strict sentence pairs, and globally deduplicated to create a highly efficient training corpus.

Dataset Details

Field	Detail
Languages	Twi (tw) ↔ English (en)
License	CC BY 4.0
Author	Mich-Seth Owusu
Organisation	Ghana NLP Community
Source dataset	ghananlpcommunity/pristine-twi-english_clean
Task	Machine Translation, Cross-lingual Transfer

How the Data Was Processed

To convert the paragraph-level source dataset into a pristine sentence-level corpus for model training, we applied a strict data processing pipeline:

1. Sentence Splitting: Both the Twi and English source texts were split using a robust regex heuristic that respects sentence-ending punctuation while guarding against false splits from common abbreviations and quoting.
2. Strict Alignment Filtering: We counted the resulting sentences for both the Twi and English sides of every row. If the number of sentences did not match perfectly (indicating a mismatched or summarized translation), the entire source row was discarded to prevent misaligned training data.
3. Standardization: Sentences were normalized, stripped of excess whitespace, and aligned into a clean 1-to-1 parallel format.
4. Global Deduplication: The dataset was hashed and globally deduplicated. Any identical sentence pairs were removed to prevent data leakage and ensure memory-efficient training.
5. Data Splitting: The final clean pairs were probabilistically shuffled into Train (90%), Validation (5%), and Test (5%) splits.

(Note: The data does not include a direction column. For bidirectional Machine Translation training, it is recommended to dynamically swap the columns during the data loading phase).

Data Structure

Each row represents a single, strictly aligned sentence pair.

Column	Type	Description
twi	string	A single Twi sentence
english	string	The corresponding fluent English translation

Usage

from datasets import load_dataset

# Load the dataset
ds = load_dataset("ghananlpcommunity/pristine-twi-english-parallel-sentences")

# View a training example
print(ds["train"][0])
# Output: {'twi': 'Mepa wo kyɛw.', 'english': 'Please.'}

Intended Uses

• Machine Translation: Training and evaluating SOTA Twi ↔ English translation models.
• Cross-lingual alignment: Building cross-lingual embeddings and sentence transformers.
• LLM Pre-training: Providing high-quality parallel supervision for training large language models to understand Akan/Twi.

Limitations & Biases

• Synthetic Origin of Translations: The English translations in the source dataset were generated by a large language model (Gemini API) and have not been manually verified sentence-by-sentence by native speakers.
• Domain constraints: The source texts are heavily grounded in Ghanaian news topics. While the corpus includes diverse styles, it may not perfectly represent all registers of spoken or deeply colloquial Twi.
• Dialectal representation: The dataset broadly represents written Twi but may lean toward standard conventions, potentially under-representing specific dialectal variations (Asante Twi, Akuapem Twi, Fante, etc.).

Citation

If you use this dataset, please cite the original Pristine Twi dataset and acknowledge the Ghana NLP Community:

@dataset{owusu2024pristinetwieng_sentences,
  author    = {Owusu, Mich-Seth and Ghana NLP Community},
  title     = {Pristine Twi–English Parallel Sentences Dataset},
  year      = {2024},
  publisher = {Hugging Face},
  url       = {[https://huggingface.co/datasets/ghananlpcommunity/pristine-twi-english-parallel-sentences](https://huggingface.co/datasets/ghananlpcommunity/pristine-twi-english-parallel-sentences)}
}

Made with ❤️ by Mich-Seth Owusu for the Ghana NLP Community.