lwagari/saiko-data

📊 Saiko Trilingual Dataset

This dataset contains 1,706,118 trilingual parallel sentence pairs engineered for training Neural Machine Translation (NMT) models. It supports bidirectional translation across Afaan Oromoo, Amharic, and English.

📖 Dataset Details

Attribute	Value
Total Pairs	1,706,118
Vocabulary Size	50,004 words
Format	Pickled Python Dictionary (.pkl)
File Size	~145 MB

🏗️ Construction Pipeline

1. Raw Sources

• MT560 (HuggingFace): Oromo-English (155k) and Amharic-English (669k).
• HornMT: Direct Oromo-Amharic News (2,030).
• Total Raw: 826,259 pairs.

2. The English Pivot Strategy

Because direct Oromo-Amharic data is extremely scarce, we used English as a mathematical bridge. We inner-joined Oromo-English and Amharic-English pairs based on identical English sentences, synthesizing 1,625,524 Oromo-Amharic pairs.

3. Conversational Oversampling

To handle everyday greetings (which were absent in the formal religious/news text), we generated 1,000+ conversational phrases and oversampled them 200 times. This forced the neural network to prioritize daily conversational fluency.

4. Data Cleaning

A 16-step regex pipeline removed HTML, URLs, emojis, and misaligned sentences. A strict Length-Ratio Filter (0.3 to 3.0) was applied to ensure high-quality alignment.

🗂️ Dataset Structure

The dataset is stored as a single lstm_trilingual_data.pkl file containing a Python dictionary with source (src), target (tgt), and the custom vocabulary (vocab).

🌍 Language Tags

The vocabulary uses special tokens to route translations:

• <2om> : Translate to Afaan Oromoo
• <2am> : Translate to Amharic
• <2en> : Translate to English

📝 Citation

If you use this dataset in your research, please cite:

@mastersthesis{wagari2026saiko, author = {Wagari, Lema}, title = {Saiko-Trilingual-LSTM: A From-Scratch NMT System for Afaan Oromoo, Amharic, and English}, school = {Jimma University}, year = {2026} }