ChartBench: A Benchmark for Complex Visual Reasoning in Charts

Introduction

We propose the challenging ChartBench to evaluate the chart recognition of MLLMs.

We improve the Acc+ metric to avoid the randomly guessing situations.

We collect a larger set of unlabeled charts to emphasize the MLLM's ability to interpret visual information without the aid of annotated data points.

Todo

• Open source: SFT internlmv2 CKPT.
• Open source: all evaluation results.
• Open source: all data of ChartBench.
• Open source: the evaluate scripts.
• Open source: the inference scripts.
• Open source: the demo data (10%).

Setup

Please follow the official repository instructions below to set up the local environment.

Inference

1. Complete the basic environment setups
2. Set prompt style for both Acc+ and NQA tasks in ./Repos/utils.py
3. Modify the default path of CKPT_PATH in ./Repos/{MODEL_NAME}/infer.py
4. Reimplement the load_model and model_gen functions
5. The results are saved in ./Result/raw/{MODEL_NAME}.jsonl by default
6. Prompt LLMs in ./Stat/gpt_filter.py to extract number values in NQA task
7. Set the parameters in ./Stat/stat_all_metric.py and the statistical results are saved in ./Stat/Paper_Table

Ranking

Citation

@article{ChartBench,
    title={ChartBench: A Benchmark for Complex Visual Reasoning in Charts},
    author={Zhengzhuo Xu and Sinan Du and Yiyan Qi and Chengjin Xu and Chun Yuan and Jian Guo},
    journal={ArXiv},
    year={2023},
    volume={abs/2312.15915},
    url={https://api.semanticscholar.org/CorpusID:266550948}
}