📊 ChartGen

Official codebase for “ChartGen: Scaling Chart Understanding via Code-Guided Synthetic Chart Generation”

ChartGen is a two-stage, fully-automated pipeline that

1. Reconstructs real-world chart images into executable Python plotting scripts with a vision-language model (Phi-3.5-Vision-Instruct).
2. Augments those scripts with a code-centric LLM (quantized Codestral-22B-v0.1), scaling a 13 K-image seed set into the ChartGen-200K dataset spanning 27 chart types and 11 plotting libraries.

✨ Key Highlights

Running both stages on 13 K seed charts yields the 222.5 K image-code pairs contained in the public ChartGen-200K dataset.

📂 Repository Layout

ChartGen/
├── chartgen/            
│   ├── chart2code.py     # stage-1 script (image → code)
│   ├── code_augment.py   # stage-2 script (code  → diversified code)
│   └── extract_plot.py   # utils: extract code & render charts from raw model output
├── README.md
├── LICENSE
└── requirements.txt      # pinned dependency versions

⚙️ Setup

# 1 Clone the repo
git clone https://github.com/<your-handle>/ChartGen.git
cd ChartGen

# 2 Create and activate an isolated environment
python -m venv .venv && source .venv/bin/activate     # macOS / Linux
# .\.venv\Scripts\Activate.ps1                        # Windows PowerShell

# 3 Install *GPU* or *CPU* wheels for PyTorch first
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118   # CUDA 11.8
# pip install torch torchvision --index-url https://download.pytorch.org/whl/cpu   # CPU-only

# 4 Install the remaining Python packages
pip install -r requirements.txt

# 5 Authenticate with Hugging Face (needed for Codestral)
huggingface-cli login     # or: export HF_TOKEN=hf_your_token

Shield:

This work is licensed under a Creative Commons Attribution 4.0 International License.