lwagari/saiko-lstm-model

🌳 Saiko-Trilingual-LSTM

A from-scratch PyTorch Neural Machine Translation (NMT) system for Afaan Oromoo ↔ Amharic ↔ English.

This model was engineered entirely from zero (no pre-trained weights) as part of a Master's thesis at Jimma University.

📖 Model Details

Attribute	Value
Architecture	Sequence-to-Sequence (Seq2Seq)
Encoder	2-Layer Bidirectional LSTM (256 hidden dim)
Attention	Bahdanau Attention (additive)
Decoder	2-Layer Unidirectional LSTM
Vocabulary	50,004 words (Custom tokenizer built from scratch)
Output Layer	Linear + Softmax (50,004 classes)
Dropout	0.1
Parameters	~326 MB checkpoint
Framework	100% Custom PyTorch

📊 Training Details

Attribute	Value
Dataset Size	1,706,118 Trilingual Pairs
Data Synthesis	English Pivot Strategy + 1,000+ oversampled conversational phrases
Optimizer	Adam (Learning Rate = 0.001)
Batch Size	64
Loss Function	Cross-Entropy Loss (ignoring <pad> tokens)
Tricks	Sutskever Reversal, Teacher Forcing (50%), Gradient Clipping (1.0)
Hardware	Kaggle T4×2 / P100 GPU

🌍 Supported Languages & Directions

• Afaan Oromoo (orm) — 8.3M+ speakers
• Amharic (amh) — 32M+ speakers
• English (eng) — Global lingua franca

6 Translation Directions handled by ONE model:

• Oromo ↔ Amharic
• Oromo ↔ English
• Amharic ↔ English

💻 Usage

from huggingface_hub import hf_hub_download
import torch

# 1. Download the weights and vocabulary
model_path = hf_hub_download(
    repo_id="lwagari/saiko-lstm-model",
    filename="lstm_trilingual.pt"
)
vocab_path = hf_hub_download(
    repo_id="lwagari/saiko-lstm-model",
    filename="vocab_trilingual.pkl"
)

# 2. Load into the custom Seq2Seq architecture 
# (See the 'streamlit_app.py' in the GitHub repo for the class definitions)