ChartMimic: Evaluating LMM’s Cross-Modal Reasoning Capability via Chart-to-Code Generation

This is the official dataset repository of ChartMimic.

1. Data Overview

ChartMimic aims at assessing the visually-grounded code generation capabilities of large multimodal models (LMMs). ChartMimic utilizes information-intensive visual charts and textual instructions as inputs, requiring LMMs to generate the corresponding code for chart rendering.

ChartMimic includes 1000 human-curated (figure, instruction, code) triplets, which represent the authentic chart use cases found in scientific papers across various domains (e.g., Physics, Computer Science, Economics, etc). These charts span 18 regular types and 4 advanced types, diversifying into 191 subcategories. Furthermore, we propose multi-level evaluation metrics to provide an automatic and thorough assessment of the output code and the rendered charts. Unlike existing code generation benchmarks, ChartMimic places emphasis on evaluating LMMs’ capacity to harmonize a blend of cognitive capabilities, encompassing visual understanding, code generation, and cross-modal reasoning.

2. Data Statistics

Type	Bar	Line	ErrorBar	Heatmap	Box	Scatters	Hist	Radar	3D	Pie	ErrorPoint	Violin
Count	100	80	30	30	25	25	20	20	15	15	10	10
Subcategories	16	8	12	4	6	4	3	6	5	8	5	3
Code Length (AVG.)	689.3	805.7	651.9	664.6	659.9	617.2	507.3	772.2	668.7	385.1	523.7	912.0
Code Length (STD.)	289.0	272.2	146.1	303.0	248.9	287.6	161.2	136.4	168.9	123.1	228.9	298.7
Level Count	53/40/7	74/6/0	17/13/0	0/19/11	15/10/0	20/5/0	13/7/0	13/7/0	2/8/5	10/5/0	7/3/0	4/6/0

Type	Area	Contour	Density	Graph	Quiver	Treemap	Combination	HR	Muiltidiff	PIP	Total
Count	5	5	5	5	5	5	30	25	25	10	500
Subcategories	2	3	4	4	4	4	30	25	25	10	101+(90)
Code Length (AVG.)	832.8	352.8	431.8	310.8	845.0	302.2	641.9	689.9	788.1	958.9	682.9
Code Length (STD.)	177.5	136.7	71.9	61.7	607.0	34.7	159.0	301.5	331.0	285.9	285.1
Level Count	4/1/0	0/0/5	3/2/0	5/0/0	0/2/3	5/0/0	3/19/8	1/4/20	0/12/13	0/0/10	249/169/82

Additional Notes: Statistics of ChartMimic. Each one in Combination, HR, Muiltidiff, and PIP is an independent subcategory. We measure code length in terms of tokens, utilizing the Llama3 tokenizer. In the level count, "A/B/C" denotes the number of chart for "easy/medium/hard" level, respectively.

3. Download Link

You can download the whole evaluation data by running the following command:

cd ChartMimic # cd to the root directory of this repository
mkdir dataset
wget https://huggingface.co/datasets/ChartMimic/ChartMimic/resolve/main/test.tar.gz
tar -xzvf test.tar.gz -C dataset

4. Data Fields

We take an instance from the Customized Mimic task as an example to illustrate the data fields of evaluation data.

{
  'Task': 'Customized Mimic',
  'ExampleID': 'bar_5',
  'Instruction': 'There is a reference image for you and I also have a set of data about average calories burned in different exercise activities. Please refer to the reference image and generate Python matplotlib code that can create a chart that exactly mirrors the style of the reference image and accurately represents my data.\n\n\n# Data for the bar chart\nlevels = [\"Beginner\", \"Intermediate\", \"Advanced\"]\nyoga = [550, 300, 350]\ncardio = [300, 450, 400]\nstrength_training = [600, 550, 600]\n\nlabels = [\"Yoga\", \"Cardio\", \"Strength Training\"]\nylabel = \"Avg. Calories Burned\"\n\n\nNow, please output the Python matplotlib code that can create a chart that exactly mirrors the style of the reference image and accurately represents my data.\nNote: You must copy the data I provide completely into your code, and ensure that the code you generate contains the data and is directly executable.\nNote that it is necessary to use figsize=(8.0, 3.0) to set the image size.\n',
  'InputFigure': 'ori_500/bar_5.png',
  'InputFigureCode': 'ori_500/bar_5.py',
  'GroundTruthFigure': 'customized_500/bar_5.png',
  'GroundTruthFigureCode': 'customized_500/bar_5.py',
  'Difficulty': 'medium',
}

Details of the data fields are as follows:

Fields	Description
Task	Either "Direct Mimic" or "Customized Mimic".
ExampleID	Composed of a combination of charttype and a number (e.g., bar_1).
Instruction	The textual description of the current test example.
InputFigure	The file name of the input image, default is "ori_500/ExampleID.png".
InputFigureCode	The file path of the Python code for the input image, default is "ori_500/ExampleID.py".
GroundTruthFigure	The file name of the ground-truth image, default is "ori_500/ExampleID.png" or "customized_500/ExampleID.png".
GroundTruthFigureCode	The file path of the Python code for the ground-truth image, default is "ori_500/ExampleID.py" or "customized_500/ExampleID.py".
Difficulty	The difficulty level, can be "easy", "medium", or "hard".

5. Citation

@article{
      shi2024chartmimic,
      title={ChartMimic: Evaluating LMM’s Cross-Modal Reasoning Capability via Chart-to-Code Generation},
      author={Chufan Shi and Cheng Yang and Yaxin Liu and Bo Shui and Junjie Wang and Mohan Jing and Linran Xu and Xinyu Zhu and Siheng Li and Yuxiang Zhang and Gongye Liu and Xiaomei Nie and Deng Cai and Yujiu Yang},
      year={2024},
      journal={arXiv preprint arXiv:2406.09961},
      url={https://arxiv.org/abs/2406.09961}
}