lucy3/aftermath_sampled_images_per_problem

Aftermath of DrawEduMath

This contains sampled_images_per_problem.json, for recreating the results of the paper titled "The Aftermath of DrawEduMath: Vision Language Models Underperform with Struggling Students and Misdiagnose Errors".

This file contains stratified samples of images per problem, with equal numbers of erroneous and error-free student images. The key is the number of images sampled. In the paper, for our experiments involving redrawn student images, we only use the samples of 1 erroneous image and 1 error-free image.

Please consult the datacard for DrawEduMath for detailed information about data source.

Quick links:

• Aftermath of DrawEduMath code: https://github.com/lucy3/aftermath_drawedumath/tree/main
• Aftermath of DrawEduMath paper: https://arxiv.org/abs/2603.00925

Example Use

import json
with open('sampled_images_per_problem.json', 'r') as infile:
    sampled_images_per_problem = json.load(infile)

License

This dataset is licensed under CC-BY-NC-4.0. It is intended for research and educational purposes following ASSISTments's Responsible Use Guidelines.

Citation

@misc{lucy2026aftermathdrawedumathvisionlanguage,
      title={The Aftermath of DrawEduMath: Vision Language Models Underperform with Struggling Students and Misdiagnose Errors}, 
      author={Li Lucy and Albert Zhang and Nathan Anderson and Ryan Knight and Kyle Lo},
      year={2026},
      eprint={2603.00925},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2603.00925}, 
}