granite-company-agent

Granite Company Agent

<div align="center"> <img src="Granite-Company-Agent-System-Flowchart.png" alt="Diagram illustrating the Granite Company Agent workflow." width="600"/> <br /> <em>Diagram illustrating the Granite Company Agent workflow.</em> </div>

Overview

The Granite Company Agent is a lightweight, fine-tuned language model built on top of a base causal language model.
It is designed to process structured company data (methodologies, courses, teachers, FAQs, testimonials) and provide accurate, context-aware responses to user queries.

Key features:

• LoRA-based parameter-efficient fine-tuning.
• Easy dataset generation and preprocessing.
• Simple inference script to chat with the model.
• Built with Hugging Face Transformers, PEFT, and PyTorch.

---

Repository Structure

granite-company-agent/
│
├── README.md
├── LICENSE
├── requirements.txt
│
├── train_granite.py       # Script to fine-tune the model
├── data_loader.py         # Utility to load CSV, XLSX, or TXT files
├── generate_data.py       # Generate synthetic company data
├── download_model.py      # Download base model from Hugging Face
├── chat_agent.py          # Simple inference/chat script

Setup

Clone the repository:

git https://huggingface.co/MateusBarros/granite-company-agent.git
cd granite-company-agent

Install dependencies:

pip install -r requirements.txt

python generate_data.py

Download the base model:

python download_model.py

Training

Train the model on your dataset:

python train_granite.py

---

Training configuration highlights:

• Epochs: 3
• Batch size: 2 (gradient accumulation used)
• Learning rate: 2e-4
• LoRA fine-tuning (low-rank adapters)
• BF16 precision The trained model and tokenizer are saved in ./granite_company_agent.

---

Inference / Chat

Use the agent interactively:

python chat_agent.py

Type your question and get responses from the fine-tuned model. Type exit or quit to end the session.

---

License

This project is released under the Apache-2.0 License.

---

References

• Hugging Face Transformers
• PEFT: Parameter-Efficient Fine-Tuning
• PyTorch