Houzeric/text-to-mermaid

🧠 mermaidGPT — Mermaid Diagram Generator from Natural Language

mermaidGPT is a small English-language model trained to generate valid Mermaid.js diagrams from natural language prompts. It allows users to create flowcharts, sequence diagrams, and more using simple instructions like “Show a user sending a request to a server.” Build with the dataset : Celiadraw/text-to-mermaid

---

🧰 Features

• English-only natural language input
• Generates Mermaid.js graph code (flowcharts, sequence diagrams, etc.)
• GPT-like architecture using PyTorch
• Lightweight and fast (trainable on a single GPU)
• CLI-based usage for generation
• Can be extended with API or frontend

---

📦 Project Structure

/
├── train.py              # Training script
├── generate.py           # Generate Mermaid code from prompt
├── tokenizer.py          # Tokenization utilities
├── model.py              # GPT-style model architecture
├── data/                 # Training data (if any)
├── examples/             # Prompt examples and output
├── checkpoints/          # Saved model checkpoints
└── README.md

---

🚀 Training

1. Build the tokenizer

python tokenizer.py

---

2. Train the model

python train.py

---

3. Hyperparameters

block_size = 128
batch_size = 32
learning_rate = 1e-3
embed_dim = 128
n_heads = 8
n_layers = 8

---

✍️ Generation Example

Prompt:

Create a flowchart showing a user sending a request to a server and receiving a response.

Output:

graph TD
    User -->|Request| Server
    Server -->|Response| User

---

📄 License

This project is released under the MIT License. See the LICENSE file for details.

---

🤗 Model on Hugging Face

Test or download the model on Hugging Face:
📦 https://huggingface.co/eric-houzelle/mermaidGPT

---

📬 Contact

For any suggestions or questions: [email protected]