LLMs_Fine_Tuning_Studio

LLMs Fine-Tuning Studio is an end-to-end Streamlit platform for fine-tuning open-source Large Language Models (LLMs) using LoRA and QLoRA techniques. The platform provides dataset management, hyperparameter configuration, training orchestration through Google Colab GPUs, experiment tracking, model storage, and interactive chat-based inference.

🧩 Current Features

✅️ Fine-tuning LLMs using LoRA and QLoRA with Unsloth and Hugging Face libraries
✅️ Dataset management using external JSON files
✅️ Modular training and inference pipelines
✅️ Configuration-driven architecture
✅️ Support for Hugging Face foundation models
✅️ Google Colab GPU training workflow
✅️ Clean and scalable project structure

🧱 Structure

LLMs_Fine_Tuning_Studio/
│
├── app.py
├── data/
├── output/
├── notebooks/
│
├── src/
│   ├── bridge/
│   ├── ui/
│   ├── training_config/
│   ├── dataset_loader.py
│   ├── model_loader.py
│   ├── train.py
│   ├── inference.py
│   └── metrics_logger.py
│
└── assets/

❯❯❯ Roadmap

☑️ Streamlit-based training dashboard
☑️ Dataset upload and management interface
☑️ Hyperparameter configuration (epochs, batch size, learning rate, etc.)
☑️ Real-time training logs and metrics visualization
☑️ LoRA and QLoRA support
☑️ Multiple model selection
☑️ Hugging Face Hub integration
☑️ Model evaluation and benchmarking
☑️ Experiment tracking and versioning

🧠 Supported LLM Models

Small LLM Models
- Llama 3.2 1B
- Llama 3.2 3B
- Qwen 2.5 1.5B
- Qwen 2.5 3B

Medium LLM Models
- Llama 3.1 8B
- Mistral 7B
- Qwen 2.5 7B
- Qwen 2.5 14B

Large LLM Models
- Qwen 2.5 32B
- Qwen 2.5 72B
- Llama 3.1 70B
- Llama 3.3 70B

🛠️ Tech Stack

✔️ Python
✔️ Unsloth
✔️ Hugging Face Transformers
✔️ TRL (Transformer Reinforcement Learning)
✔️ PEFT (LoRA / QLoRA)
✔️ PyTorch
✔️ Streamlit
✔️ Google Colab

📄 Example Dataset Format

[
  {
    "instruction": "Who is Hadi?",
    "response": "Hadi is a computer scientist specializing in AI."
  },
  {
    "instruction": "What is LLM?",
    "response": "A Large Language Model is a deep learning model trained on massive text data."
  }
]

🚧 Project Status

Active Development

The project currently supports end-to-end fine-tuning and inference workflows and is being extended into a full-featured Streamlit application for no-code LLMs fine-tuning and experimentation.

▶️ Run

pip install -r requirements.txt

streamlit run app.py

🔜 Future Work

➡️ Multi-GPU training
➡️ Automatic Colab worker execution
➡️ Hugging Face Hub model publishing
➡️ Evaluation leaderboard
➡️ RAG integration
➡️ Docker deployment
➡️ MLflow experiment tracking
➡️ Model versioning and management
➡️ Model compression

---

🧑🏻‍💻 Author

Hadi Hosseini

AI/ML Engineer | Data Engineer | Biomedical Data Scientist
➡️ www.linkedin.com/in/hadi468