UrduGPT Web App (Streamlit)

This is the official UI for UrduGPT — a custom-built English → Urdu translator powered by a Transformer-based LLM trained from scratch using PyTorch.

---

🧠 What is UrduGPT?

UrduGPT is a research and production-friendly language model built step-by-step using:

• Raw dataset from Hugging Face (English–Urdu parallel corpus)
• Byte-Pair Encoding (BPE) tokenizers trained from scratch
• Transformer architecture inspired by "Attention Is All You Need"
• PyTorch for model building & training
• Streamlit for live translation web app

---

✅ Features

• Sentence translation with beam or greedy decoding
• Token-by-token display with confidence scores
• Translation history session & export (CSV, PDF, Word)
• Local branding (logo, favicon)
• Deployable on Streamlit Cloud or Hugging Face Spaces

---

🛠️ Step-by-Step Model Building

🔹 Step 1: Load Dataset

python urdugpt_step1_dataset.py

Loads and trims parallel corpus (English–Urdu) from Hugging Face.

🔹 Step 2: Train Tokenizer

python urdugpt_step2_tokenizer.py

Trains BPE tokenizers for both source (English) and target (Urdu) languages.

🔹 Step 3: Prepare Dataloaders

python urdugpt_step2_dataloader.py

Creates PyTorch-compatible dataset & dataloaders.

🔹 Step 4–8: Transformer Model

Model code is inside urdugpt_step8_transformer.py, built from scratch:

• Embedding + Positional Encoding
• Multi-head Attention
• FeedForward + AddNorm
• Encoder and Decoder stacks
• Final Projection layer

🔹 Step 9: Train Model

python urdugpt_step9_train.py

• Uses cross-entropy loss
• Trains for N epochs and saves checkpoints (./urdugpt/model_{epoch}.pt)

🔹 Step 10: Translate (CLI)

python urdugpt_translate.py

Interactive terminal-based translation using latest model checkpoint.

🔹 Step 11: Streamlit UI

streamlit run urdugpt_web_app.py

Clean web-based frontend with history, export, and visual confidence scores.

---

📦 Installation

pip install -r requirements.txt

---

🚀 Run Locally

Make sure the model is trained and tokenizer files exist. Then run:

streamlit run urdugpt_web_app.py

---

🌐 Deploy to Streamlit Cloud

1. Clone this to a public GitHub repository
2. Go to https://streamlit.io/cloud
3. Click New App → select your repo → urdugpt_web_app.py
4. Set Python version and add requirements.txt
5. Hit Deploy 🎉

---

🤝 Open Source Plans

This project will be used:

• To demonstrate building LLMs from scratch
• As a template for multilingual translation apps
• To support fine-tuning for Urdu/Indic NLP research

We’ll invite contributors to:

• Extend to other language pairs (e.g., English → Pashto, Hindi, Bangali, Panjabi)
• Improve UI/UX (add voice input, transliteration)
• Add dataset upload & training interface

---

💡 Benefits

• Learn Transformer internals end-to-end
• Translate with your own trained model (no API needed)
• Run entirely offline or host on open platforms
• Extendable to many other NLP tasks

---

📁 Files

urdugpt_web_app.py       # Streamlit UI
urdugpt_utils.py          # Config loader (YAML)
urdugpt_step1_dataset.py
urdugpt_step2_tokenizer.py
urdugpt_step2_dataloader.py
urdugpt_step8_transformer.py
urdugpt_step9_train.py
urdugpt_translate.py
config.yaml               # All hyperparameters & paths
favicon.ico               # UI icon
urdugpt.png               # Logo for UI

---

🙌 Author

This project is proudly created and maintained by Fayaz Khan.

• 🔗 GitHub
• 💼 LinkedIn

🧾 License

MIT License. Contributions welcome.

Made with ❤️ for Urdu speakers & NLP builders.