---
license: apache-2.0
datasets:
- tatsu-lab/alpaca
language:
- en
base_model:
- distilbert/distilgpt2
pipeline_tag: text-generation
library_name: transformers
tags:
- chatbot
- instruction-tuning
- distilgpt2
- alpaca
- transformers
- fine-tuned
- offline
- flask
---
# 🧠 My Fine-Tuned Local Chatbot

A locally hosted AI chatbot powered by a fine-tuned **DistilGPT2** model using Hugging Face Transformers. This chatbot is trained on the **Stanford Alpaca Instruction Dataset**, enabling it to follow instructions and provide helpful responses — all without relying on internet access.

---

## 🚀 Features

- 🔒 **Fully local** – no internet required after setup  
- 🧠 **Fine-tuned on Stanford Alpaca-style instructions**  
- ⚡ **Fast inference** with CUDA or CPU fallback  
- 🌐 **Flask API** with simple HTML/CSS/JavaScript frontend  
- 🎨 **Customizable prompts** and response formatting  
- 🧾 **Chat history** saved using SQLite
- Dataset Used: [Stanford Alpaca](https://github.com/tatsu-lab/stanford_alpaca)

---

## 📂 Project Structure

```
├── app.py # Flask API backend
├── train.py # Script for fine-tuning the model
├── chatbot_model/
│ └── trained_model_* # Your fine-tuned model directory
├── static/
│ ├── styles.css # Frontend styles
│ └── script.js
├── templates/
│ └── index.html # Web UI
├── requirements.txt
├── README.md # You are here!
├── download.py
├── preprocess.py
├── int_db.py
├── chat_history.db # Saves history of chats
└── processed_dataset.csv
```

---

## Demo 

![image/png](https://cdn-uploads.huggingface.co/production/uploads/686acf52adf0f81a9ece24c7/vqXOBxgQKlMwSMEiatZlK.png)

---

## 💡 Sample Prompt

> **Human**: What is the capital of France?  
> **Assistant**: The capital of France is Paris.

---

## 🏋️‍♀️ Training
```
python train.py
```

This will:

Download the Stanford Alpaca dataset

Fine-tune distilgpt2

Save it inside:
chatbot_model/trained_model_YYYYMMDD_HHMMSS/

---

## 🖥️ Run the App
```
python app.py
```
Then visit: http://localhost:5005

---

## ❓ FAQ
Q: Does this work offline?
✅ Yes! Once the model is fine-tuned, no internet is needed.

Q: Can I run it on CPU?
✅ Yes, but it will be slower. A CUDA GPU is recommended for faster responses.

Q: Can I replace the model?
✅ Yes! You can fine-tune any Hugging Face-compatible model by modifying train.py.

---

## 🛠️ Tech Stack
- Flask – Web server backend
- Transformers – Hugging Face inference
- PyTorch – Deep learning engine
- HTML/CSS/JavaScript – Frontend
- Stanford Alpaca Dataset
- SQLite – For saving chat history
- Python

---

## 📜 License
MIT License – Free to use, modify, and share.

---

## 👩‍💻 Author
Jasleen Kaur Matharoo  
📧 jasleen.matharoo@s.amity.edu  
🌐 GitHub @Jasleen-05