ChartQA with Qwen3-VL

Fine-tuning Qwen3-VL-2B for chart question answering using SFT and GRPO.

Overview

This repository implements a two-stage training pipeline for chart question answering:

1. SFT (Supervised Fine-Tuning): Initial fine-tuning on ChartQA dataset
2. GRPO (Group Relative Policy Optimization): Reinforcement learning to improve format compliance and accuracy

Model

• Base Model: Qwen/Qwen3-VL-2B-Instruct
• SFT Model: Nhaass/Qwen3-VL-2B-ChartQA
• GRPO Model: Nhaass/Qwen3-VL-2B-ChartQA-GRPO

Installation

pip install -r requirements.txt

Dataset Structure

ChartQADataset/
├── train/
│   ├── train_augmented.json
│   └── png/
├── val/
│   ├── val_augmented.json
│   └── png/
└── test/
    ├── test_augmented.json
    └── png/

Training

SFT Training

cd sft
python train.py

GRPO Training

cd grpo
python train.py

Evaluation

python test_model.py

Project Structure

ChartQA/
├── sft/                  # Supervised fine-tuning
│   ├── config.py
│   ├── model.py
│   ├── data_loader.py
│   ├── collator.py
│   ├── callbacks.py
│   ├── trainer.py
│   └── train.py
│
├── grpo/                 # GRPO training
│   ├── config.py
│   ├── model.py
│   ├── data_loader.py
│   ├── rewards.py
│   ├── callbacks.py
│   ├── trainer.py
│   └── train.py
│
├── test_model.py         # Evaluation script
└── requirements.txt

Configuration

SFT Config (sft/config.py)

• Model: Qwen3-VL-2B-Thinking
• LoRA rank: 16
• Batch size: 1 (with gradient accumulation)
• Learning rate: 2e-5

GRPO Config (grpo/config.py)

• Reward components: format, accuracy, length
• Lambda weights: configurable
• KL coefficient: 0.1
• Num samples per prompt: 4

Reward Function

GRPO uses a multi-component reward:

reward = λ_format × format × (λ_acc × accuracy + λ_len × length) + λ_format × format - 1

Where:

• format: 1 if output has JSON format, 0 otherwise
• accuracy: 1 if answer matches ground truth, 0 otherwise
• length: Dynamic reward based on response length

Results

Test the model on ChartQA test set:

python test_model.py

Results are saved to test_results.json.

Demo

You can try the live demo on Hugging Face: ChartQA-Qwen3-VL-2B Demo

License

MIT License - see LICENSE file for details.

Citation

@misc{chartqa-qwen3vl,
  title={ChartQA with Qwen3-VL: SFT and GRPO Training},
  author={ChartQA Project},
  year={2026},
  url={https://github.com/yourusername/ChartQA}
}