uni9ja-intent 0.1.4

A lightweight Nigerian-focused intent classifier using SpaCy and PyTorch

Uni9ja Intent Classifier

A lightweight, high-performance Intent Classification library built on PyTorch and SpaCy. This library allows you to identify user intentions in text using word embeddings.

🛠 Features

• Three-Mode Logic: Use our pretrained model, fine-tune it with your data, or train a new one from scratch.
• SpaCy Integration: Uses en_core_web_md for high-quality semantic vectors.
• Auto-Adaptive Architecture: Automatically rebuilds the model head if you add new intent categories during fine-tuning.
• Production Ready: Includes Early Stopping, Learning Rate Scheduling, and Dropout for regularization.

---

Installation

1. Clone the repository:

git clone https://github.com/Perfect-Aimers-Enterprise/uni9ja_nlp.git
cd uni9ja_intent

2. Install the library:

pip install uni9ja-intent

3. Download the language model:

python -m spacy download en_core_web_md

---

Usage Guide

1. Using the Pretrained Model (Inference)

If you just want to use the model out of the box to predict intents:

from uni9ja_intent import IntentModel

# Initialize and load weights
model = IntentModel()
model.load_model("pretrained/uni9ja_intent.pth")

# Predict
intent, confidence = model.predict("How do I sign up for the project?")
print(f"Intent: {intent} ({confidence:.2%})")

2. Fine-tuning with New Data

Use this when you want the model to keep its old knowledge but learn a few new phrases or categories.

new_data = [
    {"text": "I want to hire a freelancer", "label": "hire_request"},
    {"text": "Is there a discount for students?", "label": "pricing_query"}
]

model.finetune(new_data, epochs=20, save_path="pretrained/uni9ja_intent_v2.pth")

3. Training from Scratch (Custom Logic)

Use this if you have a completely different dataset and want to build a brand new brain.

my_custom_data = [
    {"text": "What is the weather?", "label": "weather"},
    {"text": "Play some music", "label": "media_control"},
    # ... more data
]

# Passing data to __init__ triggers the training setup
model = IntentModel(data=my_custom_data, hidden_dims=[256, 128], dropout=0.4)
model.train(epochs=100)
model.save_model("my_custom_model.pth")

---

Project Structure

.
├── pretrained/           # Saved .pth model checkpoints
├── uni9ja_intent/        # Core Library Source
│   ├── __init__.py       # Package entry point
│   ├── model.py          # PyTorch Neural Network Architecture
│   ├── trainer.py        # IntentModel Manager Class
│   └── spacy_loader.py   # Vectorization logic
├── setup.py              # Pip installation script
└── requirements.txt      # Dependencies

---

Configuration Parameters

Parameter	Default	Description
hidden_dims	[128, 64]	The size of the hidden layers in the MLP.
dropout	0.3	Probability of dropping neurons (prevents overfitting).
threshold	0.6	Minimum confidence score to return a label.

---

Pro-Tip for Terminal Users

To see your folder structure like the one in this documentation, remember to use the command we set up earlier: tree -I '__pycache__|.git'

Would you like me to create a test.py script for you that automatically verifies if all these features are working correctly?

Loading Data in Code

You can now load your files directly using the built-in utility:

Python from uni9ja_intent import IntentModel, load_data

1. Load data from file

training_data = load_data("path/to/your_data.json")

2. Feed it to the model

model = IntentModel(data=training_data) model.train(epochs=50)