zuba 0.0.3

A lightweight Nigerian language classifier.

🇳🇬 Zuba Language Classifier

Zuba Language Classifier is a lightweight Python package for classifying text into four major Nigerian languages:

• Hausa
• Yoruba
• Igbo
• Broken English (Pidgin)

The model is built using a fine-tuned GPT-2 architecture and loads automatically from the HuggingFace Hub, so you do not need to download or bundle any model files manually.

The package supports both single-sentence and batch predictions.

---

Features

• Predicts four Nigerian languages
• Works with single strings or lists
• Auto-downloads model from HuggingFace Hub
• Fast inference and minimal dependencies

---

Installation

Install the package from PyPI:

pip install zuba

Install PyTorch if you do not already have it:

pip install torch

---

Model Setup

The classifier automatically downloads the trained GPT-2 model from the HuggingFace Hub the first time you use it.

No manual setup is required.

You may also use a custom model if your architecture differs.

---

🚀 Quick Start

from Zuba.Prediction import Classify

Classifier = Classify()

classifier=Classify()
result=classifier.predict([
    
        "Ina jin dadin wannan rana.",
        "Mo fẹ jẹ iyan.",
        "A na m aga ahịa echi."
    
])
print(result)

Output:

["hausa", "yoruba", "igbo"]

---

⚙️ How It Works

1. Your model class is loaded by the Classify constructor.
2. The package downloads the trained weights from HuggingFace automatically.
3. Text is tokenized with GPT-2 tokenizer (tiktoken).
4. Input is encoded, padded, and truncated to the required sequence length.
5. The model performs inference and outputs one of the four language labels.

---

classifier.py

Handles tokenization, preprocessing, batching, and predictions.

model_loader.py

Downloads and loads the trained model weights from HuggingFace.

---

🔧 Configuration

Tokenization Length

The default maximum token length is:

277

🔗 Important Constants

REPO_ID = "your-username/your-repo-name"
PACKAGE_LINK = "PYPI_link"

---

📜 License

This project is licensed under the MIT License.

---