lightchat 0.1.1

Lightweight GPT2 training and deployment toolkit

LightChat

LightChat is a lightweight GPT-2–based toolkit built on top of DistilGPT2. It enables anyone to train, deploy, and interact with a custom chatbot on low‑end devices using simple CLI commands.

🌐 Links & Community

• 🔗 GitHub Repository: github.com/reprompts/lightchat
• 💼 LinkedIn Group: LightChat Dev Group
• 📰 Dev.to Profile: @repromptsquest
• 🐦 Twitter: @repromptsquest

---

🔧 Features

• Train your own language model on plain text files
• Chat interactively with your fine‑tuned model
• List & delete saved models
• Supports top‑k and top‑p (nucleus) sampling

---

📋 Dataset Preparation

• Provide a plain text file (.txt) with one sentence per line.
• Aim for at least 1,000–10,000 lines for reasonable results on CPU.
• Clean, focused content yields better chat relevance.

Example (data.txt):

Hello, how can I help you today?
I love reading sci‑fi novels.
What's the weather like?

---

⚙️ Installation

pip install lightchat

⚠️ CPU install note: Transformers and PyTorch may take several minutes to compile on CPU.

---

🚀 Training

lightchat train <model_name> <data.txt> \
  --epochs 3 \
  --batch-size 8 \
  --learning-rate 5e-5

• model_name: directory under models/ to save to
• epochs: full passes over your data
• batch-size: number of samples per step
• learning-rate: step size for optimizer

⚠️ CPU training note: Training on CPU is slow. More epochs/bigger batch sizes = longer time but better fit.

---

💬 Chatting

lightchat chat <model_name> \
  --max-length 100 \
  --top-k 50 \
  --top-p 0.9 \
  --temperature 1.0

• max-length: max generated tokens per reply
• top-k: sample from top k tokens
• top-p: sample from top cumulative probability p
• temperature: randomness control (higher = more creative)

Trained models live in models/<model_name>/.

---

📂 Model Management

• List saved models: lightchat list-models
• Delete a model: lightchat delete-model <model_name>
• Or manually remove models/<model_name>/ directory.

---

🙌 Contributions

Contributions are welcome! Please see CONTRIBUTING.md.