TAIT — Termux AI Training

A lightweight, NumPy-first local AI training toolkit for Termux, Android, and constrained Linux devices.
Made in Sudan 🇸🇩

English · العربية · 简体中文

TAIT started as a single-file MiniGPT trainer and is now an open-source modular package. It stays deliberately lightweight: no PyTorch, no TensorFlow, and no large runtime stack.

✨ Features

• NumPy-only MiniGPT training and inference
• Byte-level BPE tokenizer
• Compact reasoning / CoT training format
• Response-only loss and validation loss
• Early stopping and best checkpoints
• Safe Ctrl+C checkpoint preservation
• JSONL datasets with reasoning aliases
• Arabic, English, and mixed sample datasets
• Optional Hugging Face dataset adapter
• TOML configuration with CLI overrides
• Termux-aware setup, doctor, and benchmark
• Terminal chat and local browser chat

🚀 Installation

pip install tait

From source:

git clone https://github.com/muzammilsoft/tait.git
cd tait
pip install -e .

⚡ Quick Start

From the repository:

tait data inspect examples/datasets/mixed.jsonl
tait train --config configs/reasoning.toml
tait chat --model artifacts/reasoning.npz

Browser chat:

tait chat --web --model artifacts/reasoning.npz

Show the generated compact reasoning trace:

tait chat --model artifacts/reasoning.npz --show-reasoning

📦 Dataset format

{"prompt":"شنو عاصمة السودان؟","reasoning":"السؤال عن عاصمة السودان، والمدينة المعروفة بأنها العاصمة هي الخرطوم.","response":"عاصمة السودان هي الخرطوم."}

The default dataset is examples/datasets/mixed.jsonl. Arabic and English datasets are included separately. See Datasets for detailed generation rules.

🧠 Training

tait train --config configs/default.toml

CLI options override TOML settings. Validation, early stopping, response-only loss, and compact reasoning are enabled by default in the standard configuration.

See Training.

📚 Documentation

• Architecture
• Configuration
• Datasets
• Training
• Models
• Web Chat
• Termux
• Development
• Coding Guidelines
• AI Contributions
• Release Process
• Roadmap

🗺️ Roadmap

v2.1 focuses on better instruction learning and compact reasoning. A future v2.x can add teacher/student distillation workflows. v3 can add an Agent layer with function calling, tools, planning, and execution.

License

MIT — see LICENSE.